Multi Capa

Physics Reports ( ) –
Contents lists available at ScienceDirect
Physics Reports
journal homepage: www.elsevier.com/locate/physrep
The structure and dynamics of multilayer networks

S. Boccaletti a,b,∗ , G. Bianconi c , R. Criado d,e , C.I. del Genio f,g,h ,
J. Gómez-Gardeñes i , M. Romance d,e , I. Sendiña-Nadal j,e , Z. Wang k,l ,
M. Zanin m,n
a
CNR - Institute of Complex Systems, Via Madonna del Piano, 10, 50019 Sesto Fiorentino, Florence, Italy
b
The Italian Embassy in Israel, 25 Hamered st., 68125 Tel Aviv, Israel
c
School of Mathematical Sciences, Queen Mary University of London, London, United Kingdom
d
Departamento de Matemática Aplicada, Universidad Rey Juan Carlos, 28933 Móstoles, Madrid, Spain
e
Center for Biomedical Technology, Universidad Politécnica de Madrid, 28223 Pozuelo de Alarcón, Madrid, Spain
f
Warwick Mathematics Institute, University of Warwick, Gibbet Hill Road, Coventry CV4 7AL, United Kingdom
g
Centre for Complexity Science, University of Warwick, Gibbet Hill Road, Coventry CV4 7AL, United Kingdom
h
Warwick Infectious Disease Epidemiology Research (WIDER) Centre, University of Warwick, Gibbet Hill Road, Coventry CV4 7AL,
United Kingdom
i
Institute for Biocomputation and Physics of Complex Systems, University of Zaragoza, Zaragoza, Spain
j
Complex Systems Group, Universidad Rey Juan Carlos, 28933 Móstoles, Madrid, Spain
k
Department of Physics, Hong Kong Baptist University, Kowloon Tong, Hong Kong Special Administrative Region
l
Center for Nonlinear Studies, Beijing–Hong Kong–Singapore Joint Center for Nonlinear and Complex Systems (Hong Kong) and Institute
of Computational and Theoretical Studies, Hong Kong Baptist University, Kowloon Tong, Hong Kong Special Administrative Region
m
Innaxis Foundation & Research Institute, José Ortega y Gasset 20, 28006 Madrid, Spain
n
Faculdade de Ciências e Tecnologia, Departamento de Engenharia Electrotécnica, Universidade Nova de Lisboa, 2829-516 Caparica,
Portugal
article info abstract

Article history: In the past years, network theory has successfully characterized the interaction among
Accepted 3 July 2014 the constituents of a variety of complex systems, ranging from biological to technological,
Available online xxxx and social systems. However, up until recently, attention was almost exclusively given to
editor: I. Procaccia
networks in which all components were treated on equivalent footing, while neglecting all
the extra information about the temporal- or context-related properties of the interactions
under study. Only in the last years, taking advantage of the enhanced resolution in real
data sets, network scientists have directed their interest to the multiplex character of
real-world systems, and explicitly considered the time-varying and multilayer nature
of networks. We offer here a comprehensive review on both structural and dynamical
organization of graphs made of diverse relationships (layers) between its constituents,
and cover several relevant issues, from a full redefinition of the basic structural measures,
to understanding how the multilayer nature of the network affects processes and
dynamics.
© 2014 Elsevier B.V. All rights reserved.
∗ Corresponding author.
E-mail address: stefano.boccaletti@gmail.com (S. Boccaletti).
http://dx.doi.org/10.1016/j.physrep.2014.07.001
0370-1573/© 2014 Elsevier B.V. All rights reserved.
2 S. Boccaletti et al. / Physics Reports ( ) –
Contents
1. Introduction............................................................................................................................................................................................. 3
1.1. The multilayer network approach to nature............................................................................................................................. 3
1.2. Outline of the report ................................................................................................................................................................... 5
2. The structure of multilayer networks.................................................................................................................................................... 5
2.1. Definitions and notations ........................................................................................................................................................... 6
2.1.1. The formal basic definitions........................................................................................................................................ 6
2.1.2. A small (and less formal) handbook ........................................................................................................................... 8
2.2. Characterizing the structure of multilayer networks............................................................................................................... 10
2.2.1. Centrality and ranking of nodes.................................................................................................................................. 10
2.2.2. Clustering ..................................................................................................................................................................... 12
2.2.3. Metric structures: shortest paths and distances ....................................................................................................... 15
2.2.4. Matrices and spectral properties ................................................................................................................................ 17
2.2.5. Mesoscales: motifs and modular structures .............................................................................................................. 17
2.3. Matrices and tensor representation: spectral properties......................................................................................................... 18
2.4. Correlations in multiplex networks........................................................................................................................................... 20
2.4.1. Degree correlations in multiplex networks ............................................................................................................... 21
2.4.2. Overlap and multidegree ............................................................................................................................................ 22
2.4.3. Multistrength and inverse multiparticipation ratio in weighted multiplex networks ........................................... 24
2.4.4. The activities of the nodes and the pairwise multiplexity........................................................................................ 25
3. Generative models .................................................................................................................................................................................. 27
3.1. Multiplex networks models ....................................................................................................................................................... 28
3.1.1. Growing multiplex networks models ........................................................................................................................ 28
3.1.2. Simple static models.................................................................................................................................................... 30
3.1.3. Ensembles of multiplex networks .............................................................................................................................. 30
3.2. Ensembles of networks of networks ......................................................................................................................................... 34
3.2.1. Case I: networks of networks with fixed supernetwork and links allowed only between replica nodes ............. 34
3.2.2. Case II: networks of networks with given superdegree distribution ....................................................................... 35
3.2.3. Case III: networks of networks with fixed supernetwork and random permutations of the labels of the nodes. 36
3.2.4. Case IV: networks of networks with multiple interconnections .............................................................................. 37
4. Resilience and percolation...................................................................................................................................................................... 37
4.1. The fragility of multilayer interdependent networks............................................................................................................... 38
4.2. Percolation in interdependent networks .................................................................................................................................. 39
4.2.1. The mutually connected giant component of a multilayer network ....................................................................... 39
4.3. Interdependent multiplex networks ......................................................................................................................................... 41
4.3.1. Case of a multiplex formed by M Poisson networks with the same average degree .............................................. 41
4.3.2. Case of a duplex network formed by two Poisson networks with different average degree ................................. 42
4.3.3. Multiplex networks formed by layers of scale-free networks.................................................................................. 43
4.3.4. Cascading failures ........................................................................................................................................................ 43
4.3.5. Partial interdependence .............................................................................................................................................. 46
4.3.6. Percolation in multiplex networks with multiple support-dependence relations ................................................. 47
4.4. Interdependent networks of networks ..................................................................................................................................... 48
4.4.1. Case I: networks of networks with fixed supernetwork and interdependent links allowed only between
replica nodes ................................................................................................................................................................ 48
4.4.2. Case II: networks of networks with given superdegree of the layers ...................................................................... 49
4.4.3. Case III: networks of networks with fixed supernetwork and random permutation of the labels of the nodes .. 51
4.5. Effects of correlations and embedding space on percolation properties of multilayer networks......................................... 52
4.5.1. Percolation in networks with overlap of the links .................................................................................................... 52
4.5.2. Percolation in networks with degree correlations .................................................................................................... 52
4.5.3. Percolation in spatial multiplex networks ................................................................................................................. 53
4.6. Other percolation problems ....................................................................................................................................................... 53
4.6.1. Classical percolation .................................................................................................................................................... 53
4.6.2. Percolation of antagonistic networks ......................................................................................................................... 53
4.6.3. K-core percolation ....................................................................................................................................................... 54
4.6.4. Viability in multiplex networks and weak percolation............................................................................................. 55
4.7. Cascades on multilayer networks .............................................................................................................................................. 55
5. Spreading processes and games............................................................................................................................................................. 55
5.1. Diffusion on multilayer networks.............................................................................................................................................. 56
5.1.1. Linear diffusion ............................................................................................................................................................ 56
5.1.2. Random walks.............................................................................................................................................................. 58
5.1.3. Diffusion-based centrality measures.......................................................................................................................... 60
5.1.4. Routing and congestion phenomena.......................................................................................................................... 61
5.2. Spreading processes ................................................................................................................................................................... 61
5.2.1. Disease spreading on single-layer networks ............................................................................................................. 62
5.2.2. Disease spreading on interconnected networks........................................................................................................ 64
5.2.3. Disease spreading on multiplex networks ................................................................................................................. 66
S. Boccaletti et al. / Physics Reports ( ) – 3
5.2.4. Voluntary prevention measures ................................................................................................................................. 70

5.2.5. Opinion formation dynamics ...................................................................................................................................... 72
5.3. Evolutionary games on multilayer networks............................................................................................................................ 74
5.3.1. Pairwise interaction games ......................................................................................................................................... 75
5.3.2. The public goods game ................................................................................................................................................ 77
5.3.3. Partial overlap and spatial coevolution ...................................................................................................................... 79
6. Synchronization ...................................................................................................................................................................................... 80
6.1. Synchronization in multilayer networks with alternating layers ........................................................................................... 81
6.1.1. The Master Stability Function for time dependent networks................................................................................... 81
6.1.2. Commutative layers..................................................................................................................................................... 82
6.1.3. Enhancing synchronization via evolution through commutative layers ................................................................. 84
6.1.4. Non-commutative layers ............................................................................................................................................ 86
6.1.5. Building a solution for a generic layer switching ...................................................................................................... 87
6.2. Synchronization in multilayer networks with coexisting layers............................................................................................. 94
6.2.1. Hypernetworks ............................................................................................................................................................ 94
6.2.2. Interconnected/interdependent networks and networks of networks.................................................................... 94
6.2.3. Multiplex, multiplex with time delay and master–slave configurations................................................................. 95
6.2.4. Bipartite and multilayer networks ............................................................................................................................. 95
6.2.5. Synchronization and the problem of targeting.......................................................................................................... 96
7. Applications............................................................................................................................................................................................. 98
7.1. Social............................................................................................................................................................................................ 98
7.1.1. Multilayer social networks: an old concept............................................................................................................... 98
7.1.2. Online communities .................................................................................................................................................... 99
7.1.3. Internet......................................................................................................................................................................... 102
7.1.4. Citation networks ........................................................................................................................................................ 102
7.1.5. Other social interactions ............................................................................................................................................. 102
7.2. Technical systems ....................................................................................................................................................................... 103
7.2.1. Interdependent systems.............................................................................................................................................. 103
7.2.2. Transportation systems ............................................................................................................................................... 103
7.2.3. Other technical systems .............................................................................................................................................. 105
7.3. Economy ...................................................................................................................................................................................... 105
7.3.1. Trade networks ............................................................................................................................................................ 105
7.3.2. Interbank market ......................................................................................................................................................... 106
7.3.3. Organizational networks............................................................................................................................................. 106
7.4. Other applications ...................................................................................................................................................................... 108
7.4.1. Biomedicine ................................................................................................................................................................. 108
7.4.2. Climate ......................................................................................................................................................................... 108
7.4.3. Ecology ......................................................................................................................................................................... 109
7.4.4. Psychology ................................................................................................................................................................... 109
8. Conclusions and open questions ............................................................................................................................................................ 109
Note added in proof ................................................................................................................................................................................ 112
Acknowledgments .................................................................................................................................................................................. 113
References................................................................................................................................................................................................ 113
1. Introduction
1.1. The multilayer network approach to nature
If we just turn our eyes to the immense majority of phenomena that occur around us (from those influencing our social
relationships, to those transforming the overall environment where we live, to even those affecting our own biological
functioning), we realize immediately that they are nothing but the result of the emergent dynamical organization of
systems that, on their turn, involve a multitude of basic constituents (or entities) interacting with each other via somehow
complicated patterns.
One of the major effort of modern physics is then providing proper and suitable representations of these systems, where
constituents are considered as nodes (or units) of a network, and interactions are modeled by links of that same network.
Indeed, having such a representation in one hand and the arsenal of mathematical tools for extracting information in the
other (as inherited by several gifted centuries of thoughts, concepts and activities in applied mathematics and statistical
mechanics) is the only suitable way through which we can even dare to understand the observed phenomena, identify the
rules and mechanisms that are lying behind them, and possibly control and manipulate them conveniently.
The last fifteen years have seen the birth of a movement in science, nowadays very well known under the name of complex
networks theory. It involved the interdisciplinary effort of some of our best scientists in the aim of exploiting the current
availability of big data in order to extract the ultimate and optimal representation of the underlying complex systems and
mechanisms. The main goals were (i) the extraction of unifying principles that could encompass and describe (under some
generic and universal rules) the structural accommodation that is being detected ubiquitously, and (ii) the modeling of the
resulting emergent dynamics to explain what we actually see and experience from the observation of such systems.
It would look like even pleonastic to report here on each and every original work that was carried out in specific contexts
under study (a reader, indeed, will find, along this review, all the relevant literature that was produced so far on this subject,
properly addressed and organized in the different sections of the report). At this initial stage, instead, and together with the
pioneering articles [1–4] and classical books [5–8] on complex networks, we address the interested reader to some other
reports [9–13] that were published recently on this same Journal, which we believe may constitute very good guides to
find orientation into the immensely vast literature on the subject. In particular, Ref. [9] is a complete compendium of the
ideas and concepts involved in both structural and dynamical properties of complex networks, whereas Refs. [12,13] have
the merit of accompanying and orienting the reader through the relevant literature discussing modular networks [12], and
space-embedded networks [13]. Finally, Refs. [10,11] constitute important accounts on the state of the art for what concerns
the study of synchronous organization of networking systems [11], and processes like evolutionary games on networks [10].
The traditional complex network approach to nature has mostly been concentrated to the case in which each system’s
constituent (or elementary unit) is charted into a network node, and each unit–unit interaction is represented as being a (in
general real) number quantifying the weight of the corresponding graph’s connection (or link). However, it is easy to realize
that treating all the network’s links on such an equivalent footing is too big a constraint, and may occasionally result in not
fully capturing the details present in some real-life problems, leading even to incorrect descriptions of some phenomena
that are taking place on real-world networks.
The following three examples are representative, in our opinion, of the major limitation of that approach.
The first example is borrowed from sociology. Social networks analysis is one of the most used paradigms in behavioral
sciences, as well as in economics, marketing, and industrial engineering [14], but some questions related to the real structure
of social networks have been not properly understood. A social network can be described as a set of people (or groups of
people) with some pattern of contacts or interactions between them [14,15]. At a first glance, it seems natural to assume
that all the connections or social relationships between the members of the network take place at the same level. But the
real situation is far different. The actual relationships amongst the members of a social network take place (mostly) inside
of different groups (levels or layers), and therefore they cannot be properly modeled if only the natural local-scale point of
view used in classic complex network models is taken into account. Let us for a moment think to the problem of spreading of
information, or rumors, on top of a social network like Facebook. There, all users can be seen as nodes of a graph, and all the
friendships that users create may be considered as the network’s links. However, friendships in Facebook may result from
relationships of very different origins: two Facebook’s users may share a friendship because they are colleagues in their daily
occupations, or because they are fans of the same football team, or because they occasionally met during their vacation time
in some resort, or for any other possible social reason. Now, suppose that a given user gets aware of a given information
and wants to spread it to its Facebook’s neighborhood of friends. It is evident that the user will first select that subgroup
of friends that he/she believes might be potentially interested to the specific content of the information, and only after will
proceed with spreading it to that subgroup. As a consequence, representing Facebook as a unique network of acquaintances
(and simulating there a classical diffusion model) would result in drawing incorrect conclusions and predictions of the real
dynamics of the system. Rather, the correct way to proceed is instead to chart each social group into a different layer of
interactions, and operating the spreading process separately on each layer. As we will see in Section 5, such a multilayer
approach leads actually to a series of dramatic and important consequences.
A second paradigmatic example of intrinsically multirelational (or multiplex) systems is the situation that one has to
tackle when trying to describe transportation networks, as for instance the Air Transportation Network (ATN) [16] or subway
networks [17,18]. In particular, the traditional study of ATN is by representing it as a single-layer network, where nodes
represent airports, while links stand for direct flights between two airports. On the other hand, it is clear that a more
accurate mapping is considering that each commercial airline corresponds instead to a different layer, containing all the
connections operated by the same company. Indeed, let us suppose for a moment that one wants to make predictions on the
propagation of the delays in the flight scheduling through the system, or the effects of such a dynamics on the movement of
passengers [19]. In particular, Ref. [16] considered the problem of passenger rescheduling when a fraction of the ATN links,
i.e. of flights, is randomly removed. It is well known that each commercial airline incurs in a rather high cost whenever a
passenger needs to be rescheduled on a flight of another company, and therefore it tries first to reschedule the passenger
by the use of the rest of its own network. As we will see in full details in Section 7, the proper framework where predictions
about such dynamical processes can be made suitably is, in fact, considering the ATN as a fully multilayer structure.
Moving on to biology, the third example is the effort of scientists in trying to rank the importance of a specific component
in a biological system. For instance, the Caenorhabditis elegans or C. elegans is a small nematode, and it is actually the first
ever organism for which the entire genome was sequenced. Recently, biologists were even able to get a full mapping of
the C. elegans’ neural network, which actually consists of 281 neurons and around two thousands of connections. On their
turn, neurons can be connected either by a chemical link, or by an ionic channel, and the two types of connections have
completely different dynamics. As a consequence, the only proper way to describe such a network is a multiplex graph with
281 nodes and two layers, one for the chemical synaptic links and a separate one for the gap junctions’ interactions. The most
important consequence is that each neuron can play a very different role in the two layers, and a proper ranking should be
then able to distinguish those cases in which a node of high centrality in a layer is just marginally central in the other.
These three examples, along with the many others that the reader will be presented with throughout the rest of this
report, well explain why the last years of research in network science have been characterized by more and more attempts
to generalize the traditional network theory by developing (and validating) a novel framework for the study of multilayer
networks, i.e. graphs where constituents of a system are the nodes, and several different layers of connections have to
be taken into account to accurately describe the network’s unit–unit interactions, and/or the overall system’s parallel
functioning.
Multilayer networks explicitly incorporate multiple channels of connectivity and constitute the natural environment to
describe systems interconnected through different categories of connections: each channel (relationship, activity, category)
is represented by a layer and the same node or entity may have different kinds of interactions (different set of neighbors
in each layer). For instance, in social networks, one can consider several types of different actors’ relationships: friendship,
vicinity, kinship, membership of the same cultural society, partnership or coworker-ship, etc.
Such a change of paradigm, that was termed in disparate ways (multiplex networks, networks of networks, interdepen-
dent networks, hypergraphs, and many others), led already to a series of very relevant and unexpected results, and we firmly
believe that: (i) it actually constitutes the new frontier in many areas of science, and (ii) it will rapidly expand and attract
more and more attention in the years to come, stimulating a new movement of interdisciplinary research.
Therefore, in our intentions, the present report would like to constitute a survey and a summary of the current state of
the art, together with a weighted and meditated outlook to the still open questions to be addressed in the future.
1.2. Outline of the report
Together with this introductory Preamble, the Report is organized along other 8 sections.
In the next section, we start by offering the overall mathematical definitions that will accompany the rest of our
discussion. Section 2 contains several parts that are necessarily rather formal, in order to properly introduce the quantities
(and their mathematical properties and representation) that characterize the structure of multilayer networks. The reader
will find there the attempt of defining an overall mathematical framework encompassing the different situations and
systems that will be later extensively treated. If more interested instead in the modeling or physical applications of
multilayer networks, the reader is advised, however, to take Section 2 as a vocabulary that will help and guide him/her
in the rest of the paper.
Section 3 summarizes the various (growing and not growing) models that have been introduced so far for generating
artificial multilayer networks with specific structural features. In particular, we extensively review the available literature
for the two classes of models considered so far: that of growing multiplex networks (in which the number of nodes grows in
time and fundamental rules are imposed to govern the dynamics of the network structure), and that of multiplex network
ensembles (which are ultimately ensembles of networks satisfying a certain set of structural constraints).
Section 4 discusses the concepts, ideas, and available results related to multilayer networks’ robustness and resilience,
together with the process of percolation on multilayer networks, which indeed has attracted a huge attention in recent
years. In particular, Section 4 will manifest the important differences, as far as these processes are concerned, between the
traditional approach of single-layer networks, and the case where the structure of the network has a multilayer nature.
In Section 5, we summarize the state of current understanding of spreading processes taking place on top of a multilayer
structure, and discuss explicitly linear diffusion, random walks, routing and congestion phenomena and spreading of
information and disease. The section includes also a complete account on evolutionary games taking place on a multilayer
network.
Section 6 is devoted to synchronization, and we there describe the cases of both alternating and coexisting layers. In the
former, the different layers correspond to different connectivity configurations that alternate to define a time-dependent
structure of coupling amongst a given set of dynamical units. In the latter, the different layers are simultaneously responsible
for the coupling of the network’s units, with explicit additional layer–layer interactions taken into account.
Section 7 is a large review of applications, especially those that were studied in social sciences, technology, economy,
climatology, ecology and biomedicine.
Finally, Section 8 presents our conclusive remarks and perspective ideas.
2. The structure of multilayer networks
Complexity science is the study of systems with many interdependent components, which, in turn, may interact through
many different channels. Such systems – and the self-organization and emergent phenomena they manifest – lie at the heart
of many challenges of global importance for the future of the Worldwide Knowledge Society. The development of this science
is providing radical new ways of understanding many different mechanisms and processes from physical, social, engineering,
information and biological sciences. Most complex systems include multiple subsystems and layers of connectivity, and they
are often open, value-laden, directed, multilevel, multicomponent, reconfigurable systems of systems, and placed within
unstable and changing environments. They evolve, adapt and transform through internal and external dynamic interactions
affecting the subsystems and components at both local and global scale. They are the source of very difficult scientific
challenges for observing, understanding, reconstructing and predicting their multiscale and multicomponent dynamics.
The issues posed by the multiscale modeling of both natural and artificial complex systems call for a generalization of the
‘‘traditional’’ network theory, by developing a solid foundation and the consequent new associated tools to study multilayer
and multicomponent systems in a comprehensive fashion. A lot of work has been done during the last years to understand
the structure and dynamics of these kind of systems [20–24]. Related notions, such as networks of networks [25,26],
multidimensional networks [20], multilevel networks, multiplex networks, interacting networks, interdependent networks,
and many others have been introduced, and even different mathematical approaches, based on tensor representation
[21,22] or otherwise [23,24], have been proposed. It is the purpose of this section to survey and discuss a general framework
for multilayer networks and review some attempts to extend the notions and models from single layer to multilayer
networks. As we will see, this framework includes the great majority of the different approaches addressed so far in the
literature.
2.1. Definitions and notations
2.1.1. The formal basic definitions

A multilayer networkis a pair M = (G, C ) where G = {Gα ; α ∈ {1, . . . , M }} is a family of (directed or undirected,
weighted or unweighted) graphs Gα = (Xα , Eα ) (called layers of M ) and
C = {Eαβ ⊆ Xα × Xβ ; α, β ∈ {1, . . . , M }, α ̸= β} (1)
is the set of interconnections between nodes of different layers Gα and Gβ with α ̸= β . The elements of C are called crossed
layers, and the elements of each Eα are called intralayer connections of M in contrast with the elements of each Eαβ (α ̸= β )
that are called interlayer connections.
In the remainder, we will use Greek subscripts and superscripts to denote the layer index. The set of nodes of the layer Gα
will be denoted by Xα = {xα1 , . . . , xαNα } and the adjacency matrix of each layer Gα will be denoted by A[α] = (aαij ) ∈ RNα ×Nα ,
where
if (xαi , xαj ) ∈ Eα ,

α 1
aij = (2)
0 otherwise,
αβ
for 1 ≤ i, j ≤ Nα and 1 ≤ α ≤ M. The interlayer adjacency matrix corresponding to Eαβ is the matrix A[α,β] = (aij ) ∈ RNα ×Nβ
given by:
β
if (xαi , xj ) ∈ Eαβ ,

αβ 1
aij = (3)
0 otherwise.
The projection network of M is the graph proj(M ) = (XM , EM ) where

 
 
M
 M
 M
 
XM = Xα , EM = Eα Eαβ  . (4)


α=1 α=1 α,β=1
α̸=β
We will denote the adjacency matrix of proj(M ) = (XM , EM ) by AM .

This mathematical model is well suited to describe phenomena in social systems, as well as many other complex systems.
An example is the dissemination of culture in social networks in the Axelrod Model [27], since each social group can be
understood as a layer within a multilayer network. By using this representation we simultaneously take into account:
(i) the links inside the different groups,

(ii) the nature of the links and the relationships between elements that (possibly) belong to different layers,
(iii) the specific nodes belonging to each layer involved.
A multiplex network [28] is a special type of multilayer network in which X1 = X2 = · · · = XM = X and the only possible
type of interlayer connections are those in which a given node is only connected to its counterpart nodes in the rest of layers,
i.e., Eαβ = {(x, x); x ∈ X } for every α, β ∈ {1, . . . , M }, α ̸= β . In other words, multiplex networks consist of a fixed set of
nodes connected by different types of links. The paradigm of multiplex networks is social systems, since these systems can
be seen as a superposition of a multitude of complex social networks, where nodes represent individuals and links capture
a variety of different social relations.
A given multiplex network M , can be associated to several (monolayer) networks providing valuable information about
it. A specific example is the projection network proj(M ) = (XM , EM ). Its adjacency matrix AM has elements
if aαij = 1 for some 1 ≤ α ≤ M


1
aij = (5)
0 otherwise.
A first approach to the concept of multiplex networks could suggest that these new objects are actually (monolayer)
networks with some (modular) structure in the mesoscale. It is clear that if we take a multiplex M , we can associate to it a
(monolayer) network M̃ = (X̃ , Ẽ ), where X̃ is the disjoint union of all the nodes of G1 , . . . , GM , i.e.
Xα = {xα ; x ∈ Xα }

X̃ = (6)
1≤α≤M
and Ẽ is given by
 
 
M M 
β
 
xαi , xj
α
; xαi , xjα
xαi , xi ; xi ∈ X  .
       
∈ Eα (7)


α=1 α,β=1
α̸=β
Note that M̃ is a (monolayer) graph with N × M nodes whose adjacency matrix, called supra-adjacency matrix of M , can be
written as a block matrix
 A
1 IN ··· IN

 IN A2 ··· IN 
 ..
Ã =  .. .. .. 
∈R
NM ×NM
, (8)
. . . .
IN IN ··· AM
where IN is the N-dimensional identity matrix.
The procedure of assigning a matrix to a multilayer network is often called flattening, unfolding or matricization. It is
important to remark that the behaviors of M and M̃ are related but different, since a single node of M corresponds to
different nodes in M̃ . Therefore, the properties and behavior of a multiplex M can be understood as a type of non-linear
quotient of the properties of the corresponding (monolayer) network M̃ .
It is important to remark that the concept of multilayer network extends that of other mathematical objects, such as:
1. Multiplex networks. As we stated before, a multiplex network [28] M , with M layers is a set of layers {Gα ; α ∈ {1, . . . , M }},
where each layer is a (directed or undirected, weighted or unweighted) graph Gα = (Xα , Eα ), with Xα = {x1 , . . . , xN }.
As all layers have the same nodes, this can be thought of as a multilayer network by taking X1 = · · · = XM = X and
Eαβ = {(x, x); x ∈ X } for every 1 ≤ α ̸= β ≤ M.
2. Temporal networks [29]. A temporal network (G(t ))Tt=1 can be represented as a multilayer network with a set of layers
{G1 , . . . , GT } where Gt = G(t ), Eαβ = ∅ if β ̸= α + 1, while
Eα,α+1 = {(x, x); x ∈ Xα ∩ Xα+1 } (9)
(see the schematic illustration of Fig. 1). Notice that here t is an integer, and not a continuous parameter as it will be used
later on in Section 6.1.1.
3. Interacting or interconnected networks [24]. If we consider a family of networks {G1 , . . . , GL } that interact, they can be
modeled as a multilayer network of layers {G1 , . . . , GL } and whose crossed layers Eαβ correspond to the interactions
between network Gα and Gβ (see Fig. 2).
4. Multidimensional networks [20,30–33]. Formally, an edge-labeled multigraph (multidimensional network) [30] is a triple
G = (V , E , D) where V is a set of nodes, D is a set of labels representing the dimensions, and E is a set of labeled edges,
i.e. it is a set of triples E = {(u, v, d); u, v ∈ V , d ∈ D}. It is assumed that given a pair of nodes u, v ∈ V and a label
d ∈ D, there may exist only one edge (u, v, d). Moreover, if the model considered is a directed graph, the edges (u, v, d)
and (v, u, d) are distinct. Thus, given |D| = m, each pair of nodes in G can be connected by at most m possible edges.
When needed, it is possible to consider weights, so that the edges are no longer triplets, but quadruplets (u, v, d, w),
where w is a real number representing the weight of the relation between nodes u, v ∈ V and labeled with d ∈ D. A
multidimensional network G = (V , E , D) can be modeled as a multiplex network (and therefore, as a multilayer network)
by mapping each label to a layer. Specifically, G can be associated to a multilayer network of layers {G1 , . . . , G|D| } where
for every α ∈ D, Gα = (Xα , Eα ), Xα = V ,
Eα = {(u, v) ∈ V × V ; (u, v, d) ∈ E and d = α} (10)
and Eαβ = {(x, x); x ∈ V } for every 1 ≤ α ̸= β ≤ |D|.
5. Interdependent (or layered) networks [25,34,35]. An interdependent (or layered) network is a collection of different
networks, the layers, whose nodes are interdependent to each other. In practice, nodes from one layer of the network
depend on control nodes in a different layer. In this kind of representation, the dependencies are additional edges
connecting the different layers. This structure, in between the network layers, is often called mesostructure. A preliminary
study about interdependent networks was presented in Ref. [34] (there called layered networks). Similarly to the previous
case of multidimensional networks, we can consider an interdependent (or layered) network as a multilayer network by
identifying each network with a layer.
Fig. 1. (Color online) Schematic illustration of the mapping of a temporal network into a multilayer network. (Left side) At each time t = 1, 2, 3, a different
graph characterizes the structure of interactions between the system’s constituents. (Right side) The corresponding multilayer network representation
where each time instant is mapped into a different layer.
Fig. 2. (Color online) Schematic illustration of the rather straightforward mapping of interacting networks into multilayer networks. Each different colored
network on the left side corresponds to a different blue layer on the right side.
6. Multilevel networks [36]. The formal definition of a multilevel network is the following. Let G = (X , E ) be a network. A
multilevel network is a triple M = (X , E , S ), where S = {S1 , . . . , Sp } is a family of subgraphs Sj = (Xj , Ej ), j = 1, . . . , p of
G such that
p
 p

X = Xj , E= Ej . (11)
j =1 j =1
The network G is the projection network of M and each subgraph Sj ∈ S is called a slice of the multilevel network M.
Obviously, every multilevel network M = (X , E , S ) can be understood as a multilayer network with layers {S1 , . . . , Sp }
and crossed layers Eαβ = {(x, x); x ∈ Xα ∩ Xβ } for every 1 ≤ α ̸= β ≤ p. It is straightforward to check that every
multiplex network is a multilevel network, and a multilevel network M = (X , E , S ) is a multiplex network if and only if
Xα = Xβ for all 1 ≤ α, β ≤ p.
7. Hypernetworks (or hypergraphs) [17,37–43]. A hypergraph is a pair H = (X , H ) where X is a non-empty set of nodes
and H = {H1 , . . . , Hp } is a family of non-empty subsets of X , each of them called a hyperlink of H . Now, if G = (X , E ) is
a graph, a hyperstructure S = (X , E , H ) is a triple formed by the vertex set X , the edge set E, and the hyper-edge set H.
If H = (X , H ) is a hypernetwork (or hypergraph), then it can be modeled as a multilayer network, such that for every
hyperlink h = (x1 , . . . , xk ) ∈ H we define a layer Gh which is a complete graph of nodes x1 , . . . , xk , and the crossed
layers are Eαβ = {(x, x); x ∈ Xα ∩ Xβ } (see Fig. 3).
2.1.2. A small (and less formal) handbook

A large amount of studies has shown how representing the elements of a complex system and their interactions with
nodes and links can help us to provide insights into the system’s structure, dynamics, and function. However, except for a
Fig. 3. (Color online) Schematic illustration of the transformation of a hypergraph into a multilayer network. The set of red nodes on the left side defines
three hyperlinks (H1 , H2 , and H3 ), each of which is mapped to a layer consisting of a complete graph of its nodes.
number of complex systems, the simple abstraction of their organization into a single layer of nodes and links is not sufficient.
As we have discussed in the previous section, several extensions of complex networks to multistructure or multirelational
networks have been developed in recent years [16,20–22,24,25,28,35,44–60]. In the following, we briefly describe the
characteristics of the main ones.
Amongst the mathematical models that consider non-local structures are the hypergraphs (or hypernetworks) [37] and
the hyperstructures [17]. However, these models are not able to combine the local scale with the global and the mesoscale
structures of the system. For instance, if one wants to model how a rumor spreads within a social network, it is necessary to
have in mind not only that different groups are linked through some of their members, but also that two people who know
the same person do not necessarily know each other. In fact, these two people may very well belong to entirely different
groups or levels. If one tries to model this situation with hypernetworks, then one only takes into account the social groups,
and not the actual relationships between their members. In contrast, if one uses the hyperstructure model, then one cannot
determine to which social group each contact between two nodes belongs. The key-point that makes hypernetworks and
hyperstructures not the best mathematical model for systems with mesoscale structures is that they are both node-based
models, while many real systems combine a node-based point of view with a link-based perspective.
Following on with the social network example, when one considers a relationship between two members of a social
group, one has to take into account not only the social groups that hold the members, but also the social groups that hold
the relationship itself. In other words, if there is a relationship between two people that share two distinct social groups,
such as a work and a sporting environment, one has to specify if the relationship arises in the work place, or if it has a
sport nature. A similar situation characterizes also public transport systems, where a link between two stations belonging
to several transport lines can occur as a part of different lines.
Nevertheless, the use of hypergraph analysis is far from being inadequate. In Ref. [61], the Authors investigate the theory
of random tripartite hypergraphs with given degree distributions, providing an extension of the configuration model for
traditional complex networks, and a theoretical framework in which analytical properties of these random graphs are
investigated. In particular, the conditions that allow the emergence of giant components in tripartite hypergraphs are
analyzed, along with classical percolation events in these structures. Moreover, in Ref. [62] a collection of useful metrics
on tripartite hypergraphs is provided, making this model very robust for basic analysis.
Another mathematical model capturing multiple different relations that act at the same time is that of multidimensional
networks [14,20,30–33,63–67]. In a multidimensional network, a pair of entities may be linked by different kinds of links.
For example, two people may be linked because they are acquaintances, friends, relatives or because they communicate to
each other by phone, email, or other means. Each possible type of relation between two entities is considered as a particular
dimension of the network. In the case of a multidimensional network model, a network is a labeled multigraph, that is, a
graph where both nodes and edges are labeled and where there can exist two or more edges between two nodes. Obviously,
it is possible to consider edge-only labeled multigraphs as particular models of multidimensional networks.
When edge-labeled multigraphs are used to model a multidimensional network, the set of nodes represents the set of
entities or actors in the networks, the edges represent the interactions and relations between them, and the edge labels
describe the nature of the relations, i.e. the dimensions of the network. Given the strong correlation between labels and
dimensions, often the two terms are used interchangeably. In this model, a node u belongs to (or appears in) a given
dimension d if u has at least one edge labeled with d. So, given a node u ∈ V , nu is the number of dimensions in which
u appears, i.e.
nu = |{d ∈ D; there is v ∈ V s.t. (u, v, d) ∈ E }| . (12)
Similarly, given a pair of nodes u, v ∈ V , nuv is the number of dimensions which label the edges between u and v ,
i.e. nuv = |{d ∈ D; (u, v, d) ∈ E }|. A multidimensional network can be considered as a multilayer network by making
every label equivalent to a layer.
Other alternative models focus on the study of networks with multiple kinds of interactions in the broadest as possible
sense. For example, interdependent networks were used to study the interdependence of several real world infrastructure
networks in Refs. [25,35], where the Authors also explore several properties of these structures such as cascading failures
and percolation. Related to this concept, in Ref. [35] the Authors report the presence of a critical threshold in percolation
processes, creating an analogy between interdependent networks and ideal gases. The main advantage of interdependent
networks is their ability of mapping a node in one relation with many different nodes in another relation. Thus, the
mesostructure can connect a single node in a network N1 to several different nodes in a network N2 . This is not possible
in other models where there is no explicitly defined mesostructure, and therefore a node is a single, not divisible entity.
Multilevel networks [36] lie in between the multidimensional and the interdependent networks, as they extend both
the classic complex network model and the hypergraph model [37]. Multilevel networks are completely equivalent
(isomorphic) to multidimensional networks by considering a dimension d in conjunction with the equivalent slice Sd . The
set D corresponding to dimension d is the collection of edges labeled with the label d, i.e. D = {(u, v, x); x = d}, while
Sd is the collection of nodes and edges of relation d, i.e. all the u and v that are present in at least one edge labeled with
d. In order to perform more advanced studies such as the shortest path detection, in Ref. [36] Authors introduced also a
mesostructure called auxiliary graph. Every vertex of the multilevel network M is represented by a vertex in the auxiliary
graph and, if a vertex in M belongs to two or more slice graphs in M, then it is duplicated as many times as the number
of slice graphs it belongs to. Every edge of E is an edge in the auxiliary graph and there is one more (weighted) edge for
each vertex duplication between the duplicated vertex and the original one. This operation breaks the isomorphism with
multidimensional networks and brings this representation very close to a layered network [30]. However, these two models
are not completely equivalent, since in the multilevel network the one-to-one correspondence of nodes in different slices
is strict, while this condition does not hold for layered networks. In Ref. [36], the Authors also define some extensions of
classical network measures for multilevel networks, such as the slice clustering coefficient and the efficiency, as well as a
collection of network random generators for multilevel structures.
In Ref. [68] the Authors introduce the concept of multiplex networks by extending the popular modularity function for
community detection, and by adapting its implicit null model to fit a layered network. The main idea is to represent each
layer with a slice. Each slice has an adjacency matrix describing connections between nodes belonging to the previously
considered slice. This concept also includes a mesostructure, called interslice couplings which connects a node of a specific
slice Sα to its copy in another slice Sβ . The mathematical formulation of multiplex networks has been recently developed
through many works [23,28,45,47–57,59,60]. For instance, in Ref. [23] a comprehensive formalism to deal with multiplex
systems is proposed, and a number of metrics to characterize multiplex systems with respect to node degree, edge overlap,
node participation to different layers, clustering coefficient, reachability and eigenvector centrality is provided.
Other recent extensions include multivariate networks [69], multinetworks [70,71], multislice networks [68,72–74], mul-
titype networks [75–77], multilayer networks [21,22,78,79], interacting networks [24,80,81], and networks of networks [46],
most of which can be considered particular cases of the definition of multilayer networks given in Section 2.1.1. Finally, it is
important to remark that the terminology referring to networks with multiple different relations has not yet reached a con-
sensus. In fact, different scientists from various fields still use similar terminologies to refer to different models, or distinct
names for the same model. It is thus clear that the introduction of a sharp mathematical model that fits these new structures
is crucial to properly analyze the dynamics that takes place in these complex systems which even nowadays are far from
being completely understood.
The notation proposed in Section 2.1.1 is just one of the possible ways of dealing with multilayer networks, and indeed
there have been other recent attempts to define alternative frameworks. In particular, the tensor formalism proposed in
Refs. [21,22], which we will extensively review in Section 2.3, seems promising, since it allows the synthetic and compact
expression of multiplex metrics. Nevertheless, we believe that the notation we propose here is somehow more immediate
to understand and easier to use for the study of real-world systems.
In the following, we describe the extension to the context of multilayer networks of the parameters that are traditionally
used to characterize the structural properties of a monolayer graph.
2.2. Characterizing the structure of multilayer networks
2.2.1. Centrality and ranking of nodes

The problem of identifying the nodes that play a central structural role is one of the main topics in the traditional analysis
of complex networks. In monolayer networks, there are many well-known parameters that measure the structural relevance
of each node, including the node degree, the closeness, the betweenness, eigenvector-like centralities and PageRank
centrality. In the following, we discuss the extension of these measures to multilayer networks.
One of the main centrality measures is the degree of each node: the more links a node has, the more relevant it is. The
degree of a node i ∈ X of a multiplex network M = (G, C ) is the vector [20,23]
ki = (ki , . . . , ki ),
[1] [M ]
(13)
where ki is the degree of the node i in the layer α , i.e. ki =

[α] [α]  [α]
j aij . This vector-type node degree is the natural extension
of the established definition of the node degree in a monolayer network.
One of the main goals of any centrality measure is ranking the nodes to produce an ordered list of the vertices according
to their relevance in the structure. However, since the node degree in a multiplex network is a vector, there is not a clear
ordering in RM that could produce such a ranking. In fact, one can define many complete orders in RM , and therefore we
should clarify which of these are relevant. Once one has computed the vector-type degree of the nodes, one can aggregate
this information and define the overlapping degree [23] of the node i ∈ X , as
M

ki ,
[α]
oi = (14)
α=1
i.e. oi = ∥ki ∥1 . In fact, many other aggregation measures f (ki ) could be alternatively used to compute the degree centrality,
such as a convex combination of ki , . . . , ki , or any norm of ki .
[1] [M ]
Other centrality measures, such as closeness and betweenness centrality, are based on the metric structure of the
network. These measures can be easily extended to multilayer networks, once the metric and geodesic structure are defined.
In Section 2.2.3, the reader will find a complete discussion of the metric structures in multilayer networks, which allow
straightforward extensions of this kind of centrality measures.
A different approach to measure centrality employs the spectral properties of the adjacency matrix. In particular, the
eigenvector centrality considers not only the number of links of each node but also the quality of such connections [82].
There are several different ways to extend this idea to multilayer networks, as discussed in Ref. [28], or in Ref. [83]
where eigenvector centrality is used to optimize the outcome of interacting competing networks. In the former reference,
several definitions of eigenvector-like centrality measures for multiplex networks are presented, along with studies of their
existence and uniqueness.
The simplest way to calculate eigenvector-like centralities in multiplex networks is to consider the eigenvector centrality
cα = (c1 , . . . , cN ) in each layer 1 ≤ α ≤ M separately. With this approach, for every node i ∈ X the eigenvector centrality
[α] [α]
ci is another vector
ci = (ci , . . . , ci ) ∈ RM ,
[1] [M ]
(15)
where each coordinate is the centrality in the corresponding layer. Once all the eigenvector centralities have been computed,
the independent layer eigenvector-like centrality [28] of M is the matrix
... ∈ R N ×M .
 
C = cT1 cT2 cTM (16)
Notice that C is column stochastic, since all components of cα are semipositive definite and ∥cα ∥1 = 1 for every 1 ≤ α ≤ M,
and the centrality ci of each node i ∈ X is the ith row of C . As for the degree-type indicators, a numeric centrality measure of
each node can be obtained by using an aggregation measure f (ci ) such as the sum, the maximum, or the ℓp -norm. The main
limitation of this parameter is that it does not fully consider the multilevel interactions between layers and its influence in
the centrality of each node.
If one now bears in mind that the centrality of a node must be proportional to the centrality of its neighbors (that are
distributed among all the layers), and if one considers that all the layers have the same importance, one has that
∀xαi , xαj ∈ Xα , c (xαi ) ∝ c (xαj ) if (xαj → xαi ) ∈ Gα , α ∈ {1, . . . , M }, (17)
so the uniform eigenvector-like centrality is defined [28] as the positive and normalized eigenvector
c (if it exists) of the matrix
A given by

M

A=
 (A[α] )T , (18)
α=1
where (A[α] )T is the transpose of the adjacency matrix of layer α . This situation happens, for instance, in social networks,
where different people may have different relationships with other people, while one is generically interested to measure
the centrality of the network of acquaintances.
A more complex approach is to consider different degrees of importance (or influence) in different layers of the network,
and to include this information in the definition of a matrix that defines the mutual influence between the layers. Thus, to
calculate the importance of a node within a specific layer, one must take into account also all the other layers, as some of
them may be highly relevant for the calculation. Consider, for instance, the case of a boss living in the same block of flats as
one of his employees: the relationship between the two fellows within the condominium layer formed by all the neighbors
has a totally different nature from that occurring inside the office layer, but the role of the boss (i.e. his centrality) in this
case can be even bigger than if he was the only person from the office living in that block of flats. In other words, one needs
to consider the situation where the influence amongst layers is heterogeneous.
To this purpose, one can introduce an influence matrix W = (wαβ ) ∈ RM ×M , defined as a non-negative matrix W ≥ 0
such that wαβ measures the influence on the layer Gα given by the layer Gβ . Once G and W = (wαβ ) have been fixed, the
local heterogeneous eigenvector-like centrality of G on each layer Gα is defined [28] as a positive and normalized eigenvector
cα⋆ ∈ RN (if it exists) of the matrix
M
A⋆α =

wαβ (A[β] )T . (19)
β=1
So, the local heterogeneous eigenvector-like centrality matrix can be defined as
C⋆ = c⋆1 c⋆2 ... c⋆M ∈ R N ×M .

 
(20)
A similar approach was introduced in Ref. [23], but using a matrix
M
A⋆α =

bβ (A[β] )T , (21)
β=1

where bβ ≥ 0 and β bβ = 1.
Another important issue is that, in general, the centrality of a node xαi within a specific layer α may depend not only on the
neighbors that are linked to xαi within that layer, but also on all other neighbors of xαi that belong to the other layers. Consider
the case of scientific citations in different areas of knowledge. For example, there could be two scientists working in different
subject areas (a chemist and a physicist) with one of them awarded the Nobel Prize: the importance of the other scientist will
increase even though the Nobel prize laureate had few citations within the area of the other researcher. This argument leads
to the introduction of another concept of centrality [28]. Given a multiplex network M and an influence matrix W = (wαβ ),
the global heterogeneous eigenvector-like centrality of M is defined as a positive and normalized eigenvector c ⊗ ∈ RNM (if it
exists) of the matrix
w11 (A[1] )T w12 (A[2] )T w1M (A[M ] )T

 
···
 w21 (A[1] )T w22 (A[2] )T ··· w2M (A[M ] )T 
A⊗ =  .. .. .. ..  ∈ R(NM )×(NM ) . (22)
 
 . . . . 
wL1 (A[1] )T wL2 (A[2] )T ··· wMM (A[M ] )T
Note that A⊗ is the Khatri–Rao product of the matrices
w11 w1M
 
···
 . .. .. 
W =  .. (A[1] )T (A[2] )T (A[M ] )T .
 
. .  and ··· (23)
wM1 ··· wMM
In analogy with what done before, one can introduce the notation
c⊗
 
1
 c⊗
2

c⊗ =  . , (24)
 
 .. 
c⊗
M
1 , . . . , cM ∈ R . Then, the global heterogeneous eigenvector-like centrality matrix of M is defined as the matrix given
where c⊗ ⊗ N
by
C⊗ = c⊗ c⊗ ... c⊗ ∈ R N ×M .
 
1 2 M (25)
⊗ ⊗
Note that, in general C is neither column stochastic nor row stochastic, but the sum of all the entries of C is 1.
Other spectral centrality measures include parameters based on the stationary distribution of random walkers with
additional random jumps, such as the PageRank centrality [84]. The extension of these measures and the analysis of random
walkers in multiplex networks will be discussed in Section 5.
2.2.2. Clustering
The graph clustering coefficient introduced by Watts and Strogatz in Ref. [85] can be extended to multilayer networks in
many ways. This coefficient quantifies the tendency of nodes to form triangles, following the popular saying ‘‘the friend of
your friend is my friend’’.
Recall that given a network G = (X , E ) the clustering coefficient of a given node i is defined as
# of links between the neighbors of i

cG (i) = . (26)
largest possible # of links between the neighbors of i
If we think of three people i, j and k with mutual relations between i and j as well as between i and k, the clustering coefficient
of i represents the likelihood that j and k are also related to each other. The global clustering coefficient of G is further
defined as the average of the clustering coefficients of all nodes. Obviously, the local clustering coefficient is a measure
of transitivity [86], and it can be interpreted as the density of the local node’s neighborhood.
Notice that the global clustering coefficient is sometimes defined differently, with an expression that relates it directly
to the global features of a network. This alternative definition is not equivalent to the previous one, and it is commonly used
in the social sciences [87]. Its expression, sometimes called network transitivity [14], is
# of triangles in the network

T = . (27)
# of triads in the network
In order to extend the concept of clustering to the context of multilayer networks, it is necessary to consider not only
the intralayer links, but also the interlayer links. In Ref. [36], the Authors establish some relations between the clustering
coefficient of a multilevel network, the clustering coefficient of its layers and the clustering coefficient of its projection
network. This generalization is obtained straightforwardly by identifying each slice Sq of a multilevel network M with a
layer Gα of the corresponding multiplex network M = (G, C ). Before giving a definition of the clustering coefficient of a
node i ∈ X within a multilevel network M , we need to introduce some notation.
For every node i ∈ X let N (i) be the set of all neighbors of i in the projection network proj(M ). For every α ∈ {1, . . . , M }
let Nα (i) = N (i) ∩ Xα and Sα (i) be the subgraph of the layer Gα induced by Nα (i), i.e. Sα (i) = (Nα (i), Eα (i)), where
Eα (i) = (k, j) ∈ Eα ; k, j ∈ Nα (i) .

 
(28)
Similarly, we will define S (i) as the subgraph of the projection network proj(M ) induced by N (i). In addition, the complete
graph generated by Nα (i) will be denoted by KNα (i) , and the number of links in KNα (i) by |Eα (i)|. With this notation we can
define the clustering coefficient of a given node i in M as
M
|Eα (i)|

2
α=1
CM (i) = . (29)
M
|Nα (i)|(|Nα (i)| − 1)

α=1
Then, the clustering coefficient of M can be defined as the average of all CM (i).
Once again, the clustering coefficient may be defined in several different ways. For instance, we may consider the average
of the clustering coefficients of each layer Gα . However, it is natural to opt for a definition that considers the possibility that
a given node i has two neighbors k and j with (i, k) ∈ Eα , (i, j) ∈ Eβ with α ̸= β , and (j, k) ∈ Eγ with γ ̸= α, β . This is
a situation occurring in social networks: one person i may know j from the aerobic class and k from a reading club, while j
and k know each other from the supermarket. Averaging the clustering coefficients of the layers does not help in describing
such situations, and an approach based on the projection network seems more relevant, as evidenced by the following
example: consider the multiplex network M with layers {G1 , G2 , G3 } and X = {x1 , x2 , x3 , x4 } (Fig. 4); it is easy to check that
cproj(M) (xi ) = 1 for all xi ∈ X but cGα (xi ) = 0 for all xi and α . In order to establish a relation between the clustering coefficients
of the nodes in the multiplex network, CM (i), and the clustering coefficients of the nodes in the projected network, cproj(M) (i),
we can use the same arguments employed in Ref. [36] for multilevel networks. Define θ (i) as the number of layers in which
i has less than two neighbors. Then,
1
cproj(M) (i) ≤ CM (i) ≤ cproj(M) (i). (30)
M − θ (i)
Note that Eq. (30) shows that the range of CM (i) increases with θ (i).
While the previous example (Fig. 4) shows the implicit limitations in defining the clustering coefficient of nodes in a
multiplex from that of individual layers, still it makes sense to provide an alternative ‘‘layers’’ definition.  Similarly to the
previous case, let Nα∗ (i) = {j ∈ X ; j is a neighbor of i in Gα } and Sα (i) = (Nα∗ (i), Eα (i)), where Eα (i) = (k, j) ∈ Eα ; k, j ∈
Nα∗ (i) . Note that Sα (i) is the subgraph of the layer Gα induced by Nα∗ (i). Then, the layer clustering coefficient of node i is

defined as
M
|Eα (i)|

2
ly α=1
CM (i) = . (31)
M
|Nα∗ (i)|(|Nα∗ (i)| − 1)

α=1
Notice that Sα (i) is a subgraph of Sα (i). Accordingly, Nα∗ (i) ⊆ Nα (i) and so the largest possible number of links between
neighbors of i in layer α cannot exceed the corresponding largest possible number of links between neighbors of i in Nα (i).
Fig. 4. (Color online) An example of the difference between clustering coefficients. The local clustering coefficient of all nodes is 0 in each single layer,
but 1 in the projection network.
Table 1
The adjacency matrix of the monoplex in Fig. 5.
x1 x2 x3
x1 0 1 1
x2 0 0 1
x3 1 0 0
Also, we have the relation
|N (i)|(|N (i)| − 1) M
1 
≤ |Nα∗ (i)|(|Nα∗ (i)| − 1) + |Nk∗ (i)||Nj∗ (i)|, (32)
2 2 α=1 k<j
 
terms. We can further assume that the nodes can be rearranged so that |Nα∗ (i)| ≤ |Nα+ 1 (i)| for
M ∗
where the last sum has 2
all 1 ≤ α ≤ M. Thus, using the method described in Ref. [36], we can obtain a relation between the clustering coefficient in
ly
the projected network cproj(M) (i) and the layer clustering coefficient CM (i):
θ (i)
  
ly
CM (i) ≤ M · cproj(G) (i) 1 + (M − 1) 4 + . (33)
M − θ (i)
Another possible definition of the layer clustering coefficient of a node i is the average over the clustering coefficients
cGα (i) of the slices
M
ly 1 
Cproj(M) (i) = cGα . (34)
M α=1
The following relationship between both clustering coefficients holds [36]:
|Nα∗ (i)|(|Nα∗ (i)| − 1) ly ly |Nα∗ (i)|(|Nα∗ (i)| − 1) ly

M · min Cproj(M) (i) ≤ Cproj(M) (i) ≤ M · max Cproj(M) (i). (35)
1≤α≤M 2 1≤α≤M 2
Further generalizations of the notion of clustering coefficient to multilayer networks have been proposed in Refs.
[23,47]. In Ref. [23], the Authors point out the necessity of extending the notion of triangle to take into account the richness
added by the presence of more than one layer. They define a 2-triangle as a triangle formed by an edge belonging to one
layer and two edges belonging to a second layer. Similarly, a 3-triangle is a triangle which is composed by three edges all
lying in different layers. In order to quantify the added value provided by the multiplex structure in terms of clustering, they
consider two parameters of clustering interdependence, I1 and I2 . I1 (I2 ) is the ratio between the number of triangles in the
multiplex which can be obtained only as 2-triangles (3-triangles), and the number of triangles in the aggregated system.
Then, I = I1 + I2 is the total fraction of triangles of the aggregated network which cannot be found entirely in one of the
layers. They also define a 1-triad centered at node i, for instance j − i − k, as a triad in which both edge (j − i) and edge
(i − k) are on the same layer. Similarly, a 2-triad is a triad whose two links belong to two different layers of the systems. This
way, they establish two further definitions of clustering coefficient for multiplex networks. For each node i the clustering
coefficient Ci1 is the ratio between the number of 2-triangles with a vertex in i and the number of 1-triads centered in i. A
second clustering coefficient Ci2 is defined as the ratio between the number of 3-triangles with node i as a vertex, and the
number of 2-triads centered in i. As the Authors point out, while Ci1 is a suitable definition for multiplexes with M ≥ 2, Ci2 can
only be defined for systems composed of at least three layers, and both coefficients are poorly correlated, so it is necessary to
use both clustering coefficients in order to properly quantify the abundance of triangles in multilayer networks. Averaging
over all the nodes of the system, they obtain the network clustering coefficients C1 and C2 .
In Ref. [23] the Authors also generalize the definition of transitivity. They propose two measures of transitivity: T1 as the
ratio between the number of 2-triangles and the number of 1-triads, and T2 as the ratio between the number of 3-triangles
and the number of 2-triads. As it is stressed by the Authors, clustering interdependencies I1 and I2 , average multiplex
clustering coefficients C1 and C2 , and multiplex transitivities T1 and T2 are all global network variables which give a different
perspective on the multilayer patterns of clustering and triadic closure with respect to the clustering coefficient and the
transitivity computed for each layer of the network.
In Ref. [47], the Authors derive measurements of transitivity for multiplex networks by developing several multiplex
generalizations of the clustering coefficient, and provide a comparison between some different formulations of multiplex
clustering coefficients. For instance, Authors point out that the balance between intralayer versus interlayer clustering is
different in social versus transportation networks, reflecting the fact that transitivity emerges from different mechanisms
in these cases. Such differences are rooted in the new degrees of freedom that arise from interlayer connections, and are
invisible to calculations of clustering coefficients on single-layer networks obtained via aggregation. Generalizing clustering
coefficients for multiplex networks thus makes it possible to explore such phenomena and to gain deeper insights into
different types of transitivity in networks. Further multiplex clustering coefficients are defined in Refs. [64,88,89].
2.2.3. Metric structures: shortest paths and distances

The metric structure of a complex network is related to the topological distance between nodes, written in terms of
walks and paths in the graph. So, in order to extend the classical metric concepts to the context of multilayer networks, it
is necessary to establish first the notions of path, walk and length. In order to introduce all these concepts, we will follow a
similar scheme to that used in Ref. [90]. Given a multilayer network M = (G, C ), we consider the set

E (M ) = {E1 , . . . , EM } C. (36)
A walk (of length q − 1) in M is a non-empty alternating sequence
α α α
{x1 1 , ℓ1 , x2 2 , ℓ2 , . . . , ℓq−1 , xq q }, (37)
of nodes and edges with α1 , α2 , . . . , αq ∈ {1, . . . , M }, such that for all r < q there exists an E ∈ E (M ) with
α
ℓr = xαr r , xr +r +11 ∈ E .
 
(38)
If the edges ℓ1 , ℓ2 , ℓq−1 are weighted, the length of the walk can be defined as the sum of the inverse of the corresponding
α αq
weights. If x1 1 = xq , the walk is said to be closed.
α1 α2 αq α αq
A path ω = {x1 , x2 , . . . , xq }, between two nodes x1 1 and xq in M is a walk through the nodes of M in which each node
is visited only once. A cycle is a closed path starting and ending at the same node. If it is possible to find a path between any
pair of its nodes, a multilayer network M is referred to as connected; otherwise it is called disconnected. However, different
types of reachability may be considered [20,23] depending, for example, on whether we are considering only the edges
included in some layers.
The length of a path is the number of edges of that path. Of course, in a multilayer network there are at least two types of
edges, namely intralayer and interlayer edges. Thus, this definition changes depending on whether we consider interlayer
and intralayer edges to be equivalent. Other metric definitions may be easily generalized from monolayer to multilayer
networks. So, a geodesic between two nodes u and v in M is one of the shortest path that connects u and v . The distance duv
between u and v is the length of any geodesic between u and v . The maximum distance D(M ) between any two vertices in
M is called the diameter of M . By nuv we will denote the number of different geodesics that join u and v . If x is a node and
ℓ is a link, then nuv (x) and nuv (ℓ) will denote the number of geodesics that join the nodes u and v passing through x and ℓ
respectively.
A multilayer network N = (G′ , C ′ ) is a subnetwork of M = (G, C ) if for every γ , δ with γ ̸= δ , there exist α, β with
α ̸= β such that Xγ′ ⊆ Xα , Eγ′ ⊆ Eα and Eγ′ δ ⊆ Eαβ . A connected component of M is a maximal connected subnetwork of M.
Two paths connecting the same pair of vertices in a multilayer network are said to be vertex-independent if they share no
vertices other than the starting and ending ones. A k-component is a maximal subset of the vertices such that every vertex
in the subset is connected to every other by k independent paths. For the special cases k = 2, k = 3, the k-components are
called bi-components and three-components of the multilayer network. For any given network, the k-components are nested:
every three-component is a subset of a bi-component, and so forth.
The characteristic path length is defined as
1 
L( M ) = duv , (39)
N (N − 1) u,v∈XM
u̸=v
where |XM | = N, and is also a way of measuring the performance of a graph.

The concept of efficiency, introduced by Latora and Marchiori in Ref. [91] may also be extended to multilayer networks
in a similar way to the one used in Ref. [36]. With the same notation, the efficiency of a multilayer network M is defined as
1  1
E (M ) = . (40)
N (N − 1) u,v∈XM
duv
u̸=v
It is quite natural to try to establish comparisons between the efficiency E (M ) of a multilayer network M , the efficiency
E (proj(M )) of its projection and the efficiencies of the different layers E (Gα ). In Ref. [36] there are some analytical results
that establish some relationships between these parameters.
If we consider interlayer and intralayer edges not to be equivalent, we can give alternative definitions of the metric
quantities. Let M = (G, C ) be a multilayer network, and
α α α
ω = {x1 1 , ℓ1 , x2 2 , ℓ2 , . . . , ℓq , xq+q+11 }, (41)
be a path in M . The length of ω can be defined as the non-negative value

q

ℓ(ω) = q + β ∆(j), (42)
j=2
where
if ℓj ∈ C , (i.e. if ℓj is a crossed layer)


1
∆(j) = (43)
0 otherwise,
and β is an arbitrarily chosen non-negative parameter.

In this case, the distance in M between two nodes i and j is the minimal length among all possible paths in from i to j.
Notice that for β = 0, the previous definition reduces to the natural metric in the projection network proj(M ), while
positive values of β correspond to metrics that take into account also the interplay between the different layers.
One can generalize the definition of path length even further by replacing the jumping weight β with an M × M non-
negative matrix Θ = (β(Gλ , Gµ )), to account for different distances between layers. Thus, the length of a path ω becomes
q

ℓ̃(ω) = q + ˜ (j),
∆ (44)
j =2
where
˜ (j) = β(Gσ (j−1) , Gσ (j) )


if Gσ (j) ̸= Gσ (j−1) ,
∆ (45)
0 otherwise.
The jumping weights can be further extended to include a dependence not only on the two layers involved in the layer jump,
but also on the node from which the jump starts.
For a multiplex network, it is also important to quantify the participation of single nodes to the structure of each layer
in terms of node reachability [23]. Reachability is an important feature in networked systems. In single-layer networks
it depends on the existence and length of shortest paths connecting pairs of nodes. In multilevel systems, shortest paths
may significantly differ between different layers, as well as between each layer and the aggregated topological networks.
To address this, the so-called node interdependence was introduced in Refs. [57,92]. The interdependence λi of a node i is
defined as:
 ψij
λi = , (46)
j̸=i
σij
where σij is the total number of shortest paths between node i and node j on the multiplex network, and ψij is the number of
shortest paths between node i and node j which make use of links in two or more layers. Therefore, the node interdependence
is equal to 1 when all shortest paths make use of edges laying on at least two layers, and equal to 0 when all shortest paths
use only one layer of the system. Averaging λi over all nodes, we obtain the network interdependence.
The interdependence is a genuine multiplex measure that provides information in terms of reachability. It is slightly anti-
correlated to measures of degree such as the overlapping degree. In fact, a node with high overlapping degree has a higher
number of links that can be the first step in a path toward other nodes; as such, it is likely to have a low λi . Conversely, a node
with low overlapping degree is likely to have a high value of λi , since its shortest paths are constrained to a smaller set of
edges and layers as first step. This measure is validated in Ref. [23] on the data set of Indonesian terrorists, where information
among 78 individuals are recorded with respect to mutual trust, common operations, exchanged communications and
business relationships.
2.2.4. Matrices and spectral properties

It is well known that the spectral properties of the adjacency and Laplacian matrices of a network provide insights into
its structure and dynamics [9,93,94]. A similar situation is possible for multilayer networks, if proper matrix representations
are introduced.
Given a multilayer network M , several adjacency matrices can give information about its structure: amongst others, the
most used are the adjacency matrix A[α] of each layer Gα , the adjacency matrix AM of the projection network proj(M ) and the
supra-adjacency matrix AM . The spectrum of the supra-adjacency matrix is directly related to several dynamical processes
that take place on a multilayer network. Using some results from interlacing of eigenvalues of quotients of matrices [93–95],
in Ref. [79] it was proven that if λ1 ≤ · · · ≤ λN is the spectrum of the supra-adjacency matrix AM of an undirected multilayer
network and µ1 ≤ · · · ≤ µnα is the spectrum of the adjacency matrix A[α] of layer Gα , then for every 1 ≤ k ≤ nα
λk ≤ µk ≤ λk+N −nα . (47)

A similar situation occurs for the Laplacian matrix of a multiplex network. The Laplacian matrix LM = L (also called
supra-Laplacian matrix) of a multiplex M is defined as the MN × MN matrix of the form:

...

D1β I −D12 I −D1M I
 β
D1 L1 ...
 
0 0  
...
 
 0 D2 L2 ... 0  −D21 I D2β I −D2M I 
  β .

L= . .. .. ..  + (48)

.

 . . . .   .. .. .. .. 
. . . .

... M  
0 0 DM L 
−DM1 I −DM2 I ... DM β I
 
β
In the equation above, I is the N × N identity matrix and Lα is the usual N × N Laplacian  matrix of the network layer α whose
elements are Lαij = sαi δij − wijα where sαi is the strength of node i in layer α , sαi = α
j wij . We will see in Section 5 that the
diffusion dynamics on a multiplex network is strongly related to the spectral properties of L. In Ref. [79], the Authors prove
that if λ1 ≤ · · · ≤ λN is the spectrum of the Laplacian matrix L of an undirected multilayer network and µ1 ≤ · · · ≤ µnα
is the spectrum of the Laplacian matrix Lα of layer Gα , then for every 1 ≤ k ≤ nα
µk ≤ λk+N −nα . (49)
Similar results can be found using perturbative analysis [59,96].
In addition to the spectra of the adjacency and Laplacian matrices, other types of spectral properties have been studied
for multiplex network, such as the irreducibility [97]. Several problems in network theory involve the analysis not only of
the eigenvalues, but also of the eigenvectors of a matrix [9]. This analysis typically includes the study of the existence and
uniqueness of a positive and normalized eigenvector (Perron vector), whose existence is guaranteed if the corresponding
matrix is irreducible (by using the classic Perron–Frobenius theorem). As for the spectral properties, it is possible to relate
the irreducibility of such a matrix with the irreducibility in each layer and on the projection network. In Refs. [28,97] it was
shown that if wij > 0 and the adjacency matrix AM of the projection network proj(M ) is irreducible, then the matrix A⊗
defining the global heterogeneous eigenvector-like centrality is irreducible. A similar situation occurs when we consider
random walkers in multiplex networks [98]. In this case, the uniqueness of a stationary state of the Markov process is
guaranteed by the irreducibility of a matrix of the form
 P
11 P12 ··· P1M

 P21 P22 ··· P2M 
 ∈ R(NM )×(NM ) ,
 ..
B= .. ..
.
.. (50)
. . .

PM1 PM2 ··· PMM
where Pαβ = Wαβ ⊙ A[β] + vαβ I, Wαβ ∈ RN ×N , vαβ ∈ RN , and Wαβ ⊙ A[β] is the Hadamard product of matrices Wαβ and A[β]
(see Ref. [97]). It can be proven that, under some hypotheses, if the adjacency matrix AM of the projection network proj(M )
is irreducible, then B is irreducible and hence the random walkers proposed in Ref. [98] have a unique stationary state.
2.2.5. Mesoscales: motifs and modular structures

An essential method of network analysis is the detection of mesoscopic structures known as communities (or cohesive
groups). These communities are disjoint groups (sets) of nodes which are more densely connected to each other than they
are to the rest of the network [12,99,100]. Modularity is a scalar that can be calculated for any partition of a network into
disjoint sets of nodes. Effectively, modularity is a quality function that counts intracommunity edges compared to what one
would expect at random. Thus, one tries to determine a partition that maximizes modularity to identify communities within
a network.
In Ref. [21], there is a definition of modularity for multilayer networks in which the Authors introduce a tensor that
encodes the random connections defining the null model. Ref. [68] provides a framework for the study of community
structure in a very general setting, covering networks that evolve over time, have multiple types of links (multiplexity), and
have multiple scales. The Authors generalize the determination of community structure via quality functions to multislice
networks that are defined by coupling multiple adjacency matrices.
As described above, communities are mesoscale structures mainly defined at a global level. One may ask whether more
local mesoscales, i.e. those defined slightly above the single node level, can also be extended to multilayer networks. Without
doubts, the most important of them are motifs, i.e. small sub-graphs recurring within a network with a frequency higher than
expected in random graphs [101]. Their importance resides in the fact that they can be understood as basic building blocks,
each associated with specific functions within the global system [102].
Little attention so far has been devoted to understanding the meaning of motifs in multilayer networks. An important
exception is presented in Ref. [63] in the framework of the analysis of the Pardus social network, an on-line community
that comprises 300.000 players interacting in a virtual universe. When links are grouped in two layers, corresponding to
positively and negatively connoted interactions, some resulting structures, which can be seen as multilayered motifs, appear
with a frequency much higher (e.g. triplets of users sharing positive relations) or lower (e.g. two enemy users that both have
a positive relation with a third) than expected in random networks.
2.3. Matrices and tensor representation: spectral properties
There have been some attempts in the literature for modeling multilayer networks properly by using the concept of
tensors [21,22,47]. Before describing these results, we recall some basic tensor analysis concepts.
There are two main ways to think about tensors:
(1) tensors as multilinear maps;
(2) tensors as elements of a tensor product of two or more vector spaces.
The former is more applied. The latter is more abstract but more powerful. The tensor product of two real vector spaces V
and L, denoted by V ⊗ L, consists of finite linear combinations of v ⊗ w , where v ∈ V and w ∈ L.
The dual vector space of a real vector space V is the vector space of linear functions f : V −→ R, indicated by V ∗ .
Denoting by Hom(V ∗ , L) the set of linear functions from V ∗ to L, there is a natural isomorphism between the linear
spaces V ⊗ L and Hom(V ∗ , L). Moreover, if V is a finite dimensional vector space, then there exists a natural isomorphism
(depending on the bases considered) between the linear spaces V and V ∗ . In fact, if V is finite-dimensional, the relationship
between V and V ∗ reflects in an abstract way the relation between the (1 × N ) row-vectors and the (N × 1) column-
vectors of a (N × N ) matrix. However, if V and L are finite-dimensional, then we can identify the linear spaces V ⊗ L with
Hom(V , L). Thus, a tensor σ ∈ V ⊗ L can be understood as a linear function σ : V −→ L, and therefore, once two bases
of the corresponding vector spaces have been fixed, a tensor may be identified with a specific matrix.
Now, let us consider a multiplex network M made of M layers {Gα ; 1 ≤ α ≤ M }, Gα = (Xα , Eα ), with X1 = · · · = XM =
{x1 , . . . , xN } = X , in order to describe it as a tensor product
M = ⟨X ⟩ ⊗ ⟨{ℓ1 , . . . , ℓM }⟩ , (51)
where ℓi represents the layer Gi . Linear combinations of a family of vectors are obtained by multiplying a finite number
among them by nonzero scalars (usually real numbers) and adding up the results. The span of a family of vectors is the set
of all their linear combinations. In the following, we will denote the span of a family of vectors {v1 , . . . , vm } by
span({v1 , . . . , vm }) = ⟨{v1 , . . . , vm }⟩ . (52)
Given the sets {x1 , . . . , xN } and {ℓ1 , . . . , ℓM }, we consider the formal vector spaces on R of all their linear combinations as
follows:
V = ⟨{x1 , . . . , xN }⟩ , (53)
L = ⟨{ℓ1 , . . . , ℓM }⟩ . (54)
It is straightforward to check that {x1 , . . . , xN } and {ℓ1 , . . . , ℓM } are basis of V and L respectively, and therefore
{xi ⊗ ℓα ; 1 ≤ i ≤ N , 1 ≤ α ≤ M } (55)
is a basis of V ⊗ L. So, we may give xi ⊗ ℓα the intuitive meaning of ‘‘being in the node i and in the layer ℓα ’’.
Now, if we consider a monolayer network G = (X , E ) of N nodes, with V = ⟨{x1 , . . . , xN }⟩, we can identify G with the
linear transformation σ : V −→ V such that the matrix associated to σ with respect to the basis X = {x1 , . . . , xN } is the
adjacency matrix of G.
Thus, if the adjacency matrix of G is A, we can find the set of nodes accessible from any node i in one step by multiplying
A on the left by the ith canonical basis row-vector. For example, the adjacency matrix of the graph in Fig. 5 is
0 1 1
 
0 0 1 . (56)
1 0 0
Fig. 5. Schematic illustration of the monoplex network discussed in Section 2.3 and whose adjacency matrix is given in Table 1.
Fig. 6. (Color online) Schematic illustration of a two-layer multiplex network obtained by duplicating the network of Fig. 5, and connecting each node
with its counterpart in the other layer.
Table 2
Adjacency matrix of the duplex obtained by duplicating the network of Fig. 5, and connecting each node with its counterpart
in the other layer.
x1 x2 x3
x1 ⊗ ℓ 1 x1 ⊗ ℓ 2 x2 ⊗ ℓ 1 x2 ⊗ ℓ 2 x3 ⊗ ℓ 1 x3 ⊗ℓ2
x1 x1 ⊗ ℓ1 0 1 1 0 1 0
x1 ⊗ ℓ2 1 0 0 1 0 1
x2 x2 ⊗ ℓ1 0 0 0 1 1 0
x2 ⊗ ℓ2 0 0 1 0 0 1
x3 x3 ⊗ ℓ1 1 0 0 0 0 1
x3 ⊗ ℓ2 0 1 0 0 1 0
From node 1, we can access in one step nodes 2 and 3, since
 0 1 1 0
   
= 1 .

1 0 0 0 0 1 (57)
1 0 0 1
In the same way, a multilayer network M of N nodes and with layers {ℓ1 , . . . , ℓM } can be identified with a linear
transformation σ : V ⊗ L −→ V ⊗ L where XM = {x1 , . . . , xN },
V = ⟨{x1 , . . . , xN }⟩ , (58)
L = ⟨{ℓ1 , . . . , ℓM }⟩ . (59)
It is easy to check that M is completely determined by the matrix associated to σ with respect to the basis
{xi ⊗ ℓα ; 1 ≤ i ≤ N , 1 ≤ α ≤ M }, (60)
As an example, consider the monolayer network G = (X , E ), with X = {x1 , x2 , x3 }, whose adjacency matrix A is given in
Table 1.
Now consider the multiplex network M with layers {ℓ1 , ℓ2 }, and nodes XM = {x1 , x2 , x3 }, obtained by duplicating the
previous monoplex network in both layers, and connecting each node with its counterpart in the other layer (Fig. 6). The
adjacency matrix of M is given in Table 2.
Of course, the specific matrix representing a tensor depends on the way we order the elements of the basis chosen.
{xi ⊗ ℓα ; i ∈ {1, . . . , N }, α ∈ {1, . . . , M }}. (61)
However, if A and B are two different matrices assigned to the same tensor, there exists a permutation matrix P such that
B = P · A · P −1 , and thus both matrices have the same spectral properties (see Section 2.2.4).
Fig. 7. (Color online) Schematic illustration of a generic two-layer network with interlayer connections.
Table 3
The supra-adjacency matrix (see text for definition) corresponding to the duplex network in Fig. 6.
ℓ1 ℓ2
x1 ⊗ ℓ 1 x2 ⊗ ℓ 1 x3 ⊗ ℓ1 x1 ⊗ ℓ 2 x2 ⊗ ℓ 2 x3 ⊗ℓ2
ℓ1 x1 ⊗ ℓ1 0 1 1 1 0 0
x2 ⊗ ℓ1 0 0 1 0 1 0
x3 ⊗ ℓ1 1 0 0 0 0 1
ℓ2 x1 ⊗ ℓ2 1 0 0 0 1 1
x2 ⊗ ℓ2 0 1 0 0 0 1
x3 ⊗ ℓ2 0 0 1 1 0 0
Table 4
The adjacency matrix of the two-layer network with interlayer connections sketched in Fig. 7.
x1 x2 x3
x1 ⊗ ℓ 1 x1 ⊗ ℓ 2 x2 ⊗ ℓ 1 x2 ⊗ ℓ 2 x3 ⊗ ℓ 1 x3 ⊗ℓ2
x1 x1 ⊗ ℓ1 0 1 1 1 1 0
x1 ⊗ ℓ2 1 0 0 1 0 1
x2 x2 ⊗ ℓ1 0 0 0 1 1 0
x2 ⊗ ℓ2 0 0 1 0 0 1
x3 x3 ⊗ ℓ1 1 1 0 0 0 1
x3 ⊗ ℓ2 0 1 0 0 1 0
If in the previous example we swap the order of layers and nodes, we get a different adjacency matrix, called supra-
adjacency matrix of M (Table 3), which can be written in the usual form
 
A1 I3
A= ∈ R6×6 , (62)
I3 A2
where Ai is the adjacency matrix of layer ℓi . Note that, unlike the case of multiplex networks, the blocks in the supra-
adjacency matrix of a general multilayer networks are not necessarily square matrices. It is worth mentioning the dual roles
between layers and nodes that may be observed by comparing Tables 2 and 3.
An example of a generic 2-layer network, can be seen in Fig. 7, where we can pass in one step from node x1 ⊗ ℓ1 to
node x2 ⊗ ℓ2 , and from node x3 ⊗ ℓ1 to node x1 ⊗ ℓ2 . Its adjacency matrix is listed in Table 4. It is important to stress that
while the adjacency matrix of a multilayer network with M layers and N nodes has N 2 M 2 entries, the adjacency matrix
of a multiplex network with the same layers and nodes is completely determined by only N 2 M + M 2 elements. In fact, a
multiplex network of M layers is completely determined by the M adjacency matrices corresponding to the layers and its
influence matrix W = (wαβ ) ∈ RM ×M , i.e. the non-negative matrix W ≥ 0 such that wαβ is the influence of the layer Gα on
the layer Gβ .
Multilayer networks, multidimensional networks, hypergraphs, and some other objects can be represented with tensors,
which represent a convenient formalism to implement different models [103]. Specifically, tensor-decomposition meth-
ods [103,104] and multiway data analysis [105] have been used to study various types of networks. These kind of methods
are based on representing multilayer networks as adjacency higher-order tensors and then applying generalizations of meth-
ods such as Singular Value Decomposition (SVD) [106], and the combination of CANDECOMP (canonical decomposition) by
Carroll and Chang [107] and PARAFAC (parallel factors) by Harshman [108] leading to CANDECOMP/PARAFAC (CP). These
methods have been used to extract communities [104], to rank nodes [109,110] and to perform data mining [103,111,112].
2.4. Correlations in multiplex networks
Multiplex networks encode significantly more information than their single layers taken in isolation, since they include
correlations between the role of the nodes in different layers and between statistical properties of the single layers.
Discovering the statistically significant correlations in multilayer networks is likely to be one of the major goals of network
science for the next years. Here we outline the major types of correlations explored so far. We distinguish between:
• Interlayer degree correlations
Generally speaking these correlations are able to indicate if the hubs in one layer are also the hubs, or they are typically
low degree nodes, in the other layer.
• Overlap and multidegree
The node connectivity patterns can be correlated in two or more layers and these correlations can be captured by the
overlap of the links. For example, we usually have a large fraction of friends with which we communicate through
multiple means of communications, such as phone, email and instant messaging. This implies that the mobile phone
social network has a significant overlap with the one of email communication or the one of instant messaging. The overlap
of the links can be quantified by the global or local overlap between two layers, or by the multidegrees of the nodes that
determine the specific overlapping pattern.
• Multistrengths and inverse multiparticipation ratio of weighted multiplex
The weights of the links in the different layers can be correlated with other structural properties of the multiplex.
For example, we tend to cite collaborators differently from other scientists. These types of correlations between
structural properties of the multiplex and the weights distribution are captured by the multistrengths and inverse
multiparticipation ratio.
• Node pairwise multiplexity
When the nodes are not all active in all layers two nodes can have correlated activity patterns. For example they can be
active on the same, or on different layers. These correlations are captured by the Node Pairwise Multiplexity.
• Layer pairwise multiplexity
When the nodes are not all active in all layers, two layers can have correlated activity patterns. For example they
can contain the same active nodes, or different active nodes. These correlations are captured by the Layer Pairwise
Multiplexity.
2.4.1. Degree correlations in multiplex networks

Every node i = 1, 2, . . . , N of a multiplex has a degree ki in each layer α = 1, 2, . . . , M. The degrees of the same node
[α]
in different layers can be correlated. For example, a node that is a hub in the scientific collaboration network is likely to
be a hub also in the citation network between scientists. Therefore, the degree in the collaboration network is positively
correlated with that of the citation network. Negative correlations also exist, when the hubs of one layer are not the hubs of
another layer. The methods to evaluate the degree correlations between a layer α and a layer β are the following:
• Full characterization of the matrix P (kα , kβ ).
One can construct the matrix
N (kα , kβ )
P (kα , kβ ) = , (63)
N
where N (kα , kβ ) is the number of nodes that have degree kα in layer α and degree kβ in layer β . From this matrix, the
full pattern of correlations can be determined.
• Average degree in layer α conditioned on the degree of the node in layer β
A more coarse-grained measure of correlation is k¯α (kβ ), i.e. the average degree of a node in layer α conditioned to the
degree of the same node in layer β :
kα P (kα , kβ )

α β α α β kα

k̄ (k ) = k P (k |k ) = . (64)
P (kα , kβ )

kα
kα
If this function does not depend on kβ , the degrees in the two layers are uncorrelated. If this function is increasing
(decreasing) in kβ , the degrees of the nodes in the two layers are positively (negatively) correlated.
• Pearson, Spearman and Kendall correlations coefficients
Even more coarse-grained correlation measures are the Pearson, the Spearman and the Kendall’s correlations
coefficients.
The Pearson correlation coefficient rαβ is given by
    
[α] [β] [β]
ki ki − k[α]
i ki ,
rαβ = (65)
σα σβ
   2
where σα =
[α] [α]
ki ki − k[α]
i . The Pearson correlation coefficient can be dominated by the correlations of the high
degree nodes if the degree distributions are broad.
[β]
The Spearman correlation coefficient ραβ between the degree sequences {ki } and {ki } in the two layers α and β is
[α]
given by
β    β
 
xαi xi − xαi xi
ραβ = β
, (66)
σ (xα )σ (xi )
β
where xαi is the rank of the degree ki
[α] [α] [β] [β]
in the sequence {ki } and xi is the rank of the degree ki in the sequence {ki }.
   2
β β β
  2 
Moreover, σ (xαi ) = (xαi )2 − xαi and σ (xi ) = (xi )2 − xi . The Spearman coefficient has the problem that the
ranks of the degrees in each degree sequence are not uniquely defined because degree sequences must always have some
degeneracy. Therefore, the Spearman correlation coefficient is not a uniquely defined number.
[β]
The Kendall’s τ correlation coefficient between the degree sequences {ki } and {ki } is a measure that takes into
[α]
β
account the sequence of ranks {xαi } and {xi }. A pair of nodes i and j is concordant if their ranks have the same order in the
β β
two sequences, i.e. (xαi − xαj )(xi − xj ) > 0, and discordant otherwise. The Kendall’s τ is defined in terms of the number
nc of concordant pairs and the number nd of discordant pairs and is given by
nc − nd
τ= √ . (67)
(n0 − n1 )(n0 − n2 )
In the equation above, n0 = 1/2N (N − 1) and the terms n1 and n2 account for the degeneracy of the ranks and are given
by
1
n1 = un (un − 1),
2 n
1
n2 = vn (vn − 1), (68)
2 n
where un is the number of nodes in the nth tied group of the degree sequence {ki }, and vn is the number of nodes in the
[α]
[β]
nth tied group of the degree sequence {ki }.
The level of degree correlations in the multiplex networks can be tuned by reassigning new labels to the nodes of the
[1] [2]
multiplexes [48]. Let us for simplicity consider a duplex where we have two degree sequences {ki } and {ki } as proposed
in Ref. [48]. The label of the nodes in layer 1 and layer 2 can be reassigned in order to generate the desired correlations
between the two degree sequences. Assume we have ranked the degree sequences in the two layers. If the labels of the
nodes in both sequences are given according to the increasing rank of the degree in each layer, then the sequences are
maximally-positive correlated (MP), i.e. the nodes of equal rank in the two sequences are replica nodes of the multiplex.
Conversely, if the label of a node in one layer is given according to the increasing rank of the degree in the same layer, and
the label of a node in the other layer is given according to the decreasing rank of the degree in the same layer, the two degree
sequences are maximally-negative correlated (MN). In Fig. 8, an example is given for the construction of maximally-positive
(MP) correlations and maximally-negative (MN) correlations from two given degree sequence in two layers of a multiplex
network.
2.4.2. Overlap and multidegree

A large variety of data sets, including the multiplex airport network, on-line social games, collaboration and citation
networks [63,78,113], display a significant overlap of the links. This means that the number of links present at the same
time in two different layers is not negligible with respect to the total number of links in the two layers.
The total overlap Oαβ between two layers α and β [49,63] is defined as the total number of links that are in common
between layer α and layer β :
β
Oαβ = aαij aij ,

(69)
i<j
where α ̸= β .
αβ
It is sometimes useful also to define the local overlap oi between two layers α and β [49], defined as the total number
of neighbors of node i that are neighbors in both layer α and layer β :
N
αβ β
aαij aij .

oi = (70)
j =1
The total and local overlap in multilayer networks can be very significant. Take for example the online social game studied
in Ref. [63]. In this data set the friendship and the communication layers have a significant overlap, similarly to the
MP
UC
MN
Fig. 8. (Color online) Schematic illustration of three kinds of correlated multiplex networks, maximally-positive (MP), uncorrelated (UC), and maximally-
negative (MN). Each layer of the networks has different types of links, indicated by solid and dashed links, respectively.
Source: Reprinted figure with permission from Ref. [48].
© 2014, by the American Physical Society.
Fig. 9. (Color online) Example of all possible multilinks in a multiplex network with M = 2 layers and N = 5 nodes. Nodes i and j are linked by one
⃗ = (mα , mα′ ).
multilink m
Source: Reprinted figure from Ref. [113].
communication and trade layers. Even the two ‘‘negative’’ layers of enmity and attack have significant overlap of the links.
As a second example of multiplex network with significant overlap, consider the APS data set of citations and collaboration
networks [113]. The two layers in this data set display significant overlap because two co-authors are also usually citing
each other in their papers.
One way to characterize the link overlap is by introducing the concept of multilinks [49,113]. A multilink fully determines
all the links present between any given two nodes i and j in the multiplex. Consider for example the multiplex with M = 2
layers, i.e. the duplex shown in Fig. 9. Nodes 1 and 2 are connected by one link in the first layer and one link in the second.
Thus, we say that the nodes are connected by a multilink (1, 1). Similarly, nodes 2 and 3 are connected by one link in the
first layer and no link in layer 2. Therefore, they are connected by a multilink (1, 0). In general, for a multiplex of M layers
we can define multilinks, and multidegrees in the following way. Let us consider the vector m ⃗ = (m1 , m2 , . . . , mα , . . . , mM )
in which every element mα can take only two values mα = {0, 1}. We define a multilink m ⃗ as the set of links connecting
a given pair of nodes in the different layers of the multiplex: mα = 1 if and only if they are connected in layer α . We can
⃗
therefore introduce the multiadjacency matrices Am⃗ with elements Am ⃗ between node i and
ij equal to 1 if there is a multilink m
node j, and zero otherwise:
M
aαij mα + (1 − aαij )(1 − mα ) .

⃗
Am
 
ij = (71)
α=1
Thus, multiadjacency matrices satisfy the condition


ij = 1,
⃗
Am (72)
⃗
m
for every fixed pair of nodes (i, j). Multiadjacency matrices are a way to describe the same information encoded in the M
adjacency matrices of each layer, and while they do not encode more information, they can be useful tools to characterize
⃗
partial overlap and correlations. Finally, we define the multidegree m⃗ of a node i, km ⃗
i as the total number of multilinks m
incident to node i, i.e.
N

ij .
⃗ ⃗
km
i = Am (73)
j =1
⃗ have the explicit expression in terms of the adjacency

⃗ ̸= 0,
For a duplex the non trivial multidegrees, i.e. the multidegrees m
matrix a1 of layer 1 and a2 of layer 2 given by
N
(1,0)

ki = a1ij (1 − a2ij ),
j =1
N
(0,1)

ki = (1 − a1ij )a2ij ,
j =1
N
(1,1)

ki = a1ij a2ij . (74)
j =1
Considering the example of a social duplex, formed by the same set of agents interacting by mobile phone in layer 1 and by
(1,0) (0,1)
email in layer 2, ki indicates the number of acquaintances of node i that only communicate with it via mobile phone, ki
(1,1)
indicates the number of acquaintances of node i that only communicate with it via email, and ki indicates the number of
acquaintances of node i that communicate with it via both mobile phone and email.
When the number of layers is very large, it is difficult to keep track of all the different types of multilinks. In this case one
can consider the multiplicity of the overlap νij between node i and node j indicating the total number of layers in which the
two nodes are connected, i.e.
M M
aαij = mαij ,
 
νij = (75)
α=1 α=1
⃗ =m
where the nodes i and j are linked by a multilink m ⃗ ij .
2.4.3. Multistrength and inverse multiparticipation ratio in weighted multiplex networks

In weighted multiplex networks the weights might be correlated with the multiplex network structure in a non trivial
way. To study weighted multiplex networks, two new measures, the multistrengths and the inverse multiparticipation ratio
have been introduced in Ref. [113]. For layer α associated to multilinks m, ⃗ such that mα > 0, we define the multistrength
⃗ ⃗
sm
i,α and the inverse multiparticipation ratio Y m
i,α of node i, as
N
aαij Am

ij ,
⃗ ⃗
sm
i,α = (76)
j =1
 2
N α m⃗
 a A
 ij ij 
  α m⃗  ,
⃗
Yim
,α = (77)
j =1 air Air
r
i,α measures the total weights of the links incident to node i in layer α that form a multilink
⃗
respectively. The multistrength sm
⃗ For example, in a social multiplex network where each layer is either mobile phone communication (layer 1) or
of type m.
(1,0)
email communication (layer 2), si,1 is the total strength of phone calls of node i with people that do not communicate with
(0,1)
it by email, si,2 is the total strength of email contacts of node i with people that do not communicate with it via mobile
(1,1) (1,1)
phone, si,1 is the total strength of phone calls of node i with people that also communicate with it via email, and si,2 is the
total strength of email communication of node i with people that also communicate with it via mobile phone. The inverse
⃗
multiparticipation ratio Yim ,α is a measure of the inhomogeneity of the weights of the nodes that are incident to node i in
layer α and at the same time are part of a multilink m.⃗ Since for every layer α and node i the non trivial multistrengths must
⃗
involve multilinks m ⃗ with mα = 1, the number of non trivial multistrengths sm i,α is given by 2
M −1
. Therefore, for each node,
the number of multistrengths that can be defined in a multiplex network of M layers is M2M −1 . Similarly to what happens
 for
⃗ ,α
single layers [114–116], the average multistrength of nodes with a given multidegree, i.e. sm⃗ ,α (km⃗ ) = si δ(kmi , k ) , and
⃗ ⃗
m m
 
⃗ ,α
the average inverse multiparticipation ratio of nodes with a given multidegree, Y m⃗ ,α (km⃗ ) = Yi δ(kmi , k ) , are expected
m ⃗ ⃗
m
to scale as
sm⃗ ,α (km⃗ ) ∝ (km⃗ )βm⃗ ,α
Y m⃗ ,α (km⃗ ) ∝ ⃗ )λm
⃗ ,α
1
, (78)
(km
with exponents βm⃗ ,α ≥ 1 and λm⃗ ,α ≤ 1. In Ref. [113] the weighted multilink properties of multiplex networks between
scientist collaborating (in layer 1) and citing each other (in layer 2) have been analyzed starting from the APS data set. In this
case the citation network is also directed, and the Authors find that the exponents λ and β depend on the type of multilink,
i.e. they are significantly different for links without overlap and for links with overlap. It is clear from this analysis that many
weighted multiplex networks have a distribution of the weights that depends on the non-trivial correlations existing in their
structure. In Fig. 10, the multistrength and inverse multiparticipation ratio for the collaboration and citation networks of
PRE Authors are shown.
2.4.4. The activities of the nodes and the pairwise multiplexity

From the M adjacency matrices aα of each layer α = 1, 2 . . . M of the network it is possible to construct an N × M activity
matrix b of elements bi,α indicating if node i is present in layer α . This matrix can be seen as an adjacency matrix between
nodes and layers indicating which node is active in each layer of the multiplex. For a undirected multiplex network, node i
is active in layer α if it has a positive degree in layer α , i.e. ki > 0. Therefore we have
[α]
bi,α = 1 − δ0,k[α] = 1 − δ 
N , (79)
i 0, aα
ij
j=1
where δx,y indicates the Kronecker delta. For a directed multiplex network, node i is inactive in layer α if both its in-degree
and its out-degree in layer α are zero. Therefore we can define the matrix elements bi,α as
bi,α = 1 − δ 
N δ N . (80)
aα
 α
0, ij
0, aji
j=1 j=1
The activity Bi of a node i has been defined in Ref. [117] and is given by the number of layers in which node i is active:
M

Bi = bi,α . (81)
α=1
The layer activity Nα has been defined in Ref. [117] and is given by the number of nodes active in layer α :
N

Nα = bi,α . (82)
i=1
By analyzing a large set of multiplex networks, including a multiplex formed by a very large number of layers, Nicosia
and Latora in Ref. [117] have shown that the activity distribution P (Bi ) is typically broad and can be fitted by a power-law
P (Bi ) ≃ B−δ
i with δ ∈ [1.5, 3.0]. This implies that for some multiplex networks the bipartite network between nodes and
layers described by the activity adjacency matrix can be either dense (δ ≤ 2) or sparse (δ > 2) but the typical number
of layers in which a node is active is always subject to unbound fluctuations (see Fig. 11). Moreover, in Ref. [117] a broad
distribution P (Nα ) of layer activities has been reported.
In Ref. [117] the Authors studied the layer pairwise multiplexity Qαβ measuring the correlation between the layers. The
layer pairwise multiplexity is defined as
N
1 
Qαβ = bi,α bi,β , (83)
N i =1
PRE PRE PRE

collaborations citations in citations out
s(m1,m2)
Y(m1,m2)
k(m1,m2) k(m1,m2) k(m1,m2)
Fig. 10. (Color online) Properties of multilinks in the weighted collaboration (layer 1)/citation (layer 2) multiplex network based on the PRE publications
analyzed in Ref. [113]. In the case of the collaboration network, the distributions of multistrengths versus multidegrees always have the same exponent,
but the average weight of multilinks (1, 1) is larger than the average weight of multilinks (1, 0). Moreover, the exponents λ(1,0),col,in , λ(1,0),col,out are larger
than exponents λ(1,1),col,in , λ(1,1),col,out . In the case of the citation layer, both the incoming multistrengths and the outgoing multistrengths have a functional
behavior that varies depending on the type of multilink. Conversely, the average inverse multiparticipation ratio in the citation layer does not show any
significant change of behavior when compared across different multilinks.
Fig. 11. (Color online) Distributions P (Bi ) of the node-activity Bi for (a) the six continental airline multiplex networks and for (b) APS and IMDb. Note that
these distributions are broad and can be fitted by power-law exponents between 1.5 and 3.0 indicating that the typical number of layers in which a node
is active is subject to unbound fluctuations.
Source: Reprinted figure from Ref. [117]. Courtesy of V. Latora.
Fig. 12. The network between 20 European airline companies in which each link is weighted with their layer pairwise multiplexity. This method can
reveal non trivial correlations between the activities of the nodes in the different layers.
Source: Reprinted figure from Ref. [117]. Courtesy of V. Latora.
and quantifies the fraction of nodes that are active in layer α as well as in layer β . Similarly, one can use the node pairwise
multiplexity Qij , measuring the correlation of activities between two nodes. The node pairwise multiplexity, introduced in
Ref. [36], is defined as
M
1 
Qij = bi,α bi,β , (84)
M α=1
and quantifies the fraction of layers in which both node i and node j are active.
In Fig. 12 the network of European airline companies is plotted, where the links between two different companies is
weighted with the layer pairwise multiplexity. This method can reveal non trivial correlations between the activities of the
nodes in different layers.
3. Generative models
In the last fifteen years large attention has been devoted to unveil the basic mechanisms for the generation of monolayer
networks with specific structural properties. The discussion of all generative models for single-layer networks is beyond
the scope of this report, and we refer the interested reader to the existing reviews [2–4,9,118] and to more recent works
[119,120] on the subject. Here we mention only that the models introduced so far in the Literature can be cast in two major
categories:
• the growing networks models, including the mechanism of preferential attachment introduced in the Barabási–Albert (BA)
model to explain the emergence of power-law distributions, and
• the static models, describing networks ensembles with given structural properties. Examples of such structural properties
are: the degree sequence (the configuration model), the expected degree sequence (the exponential random graph,
or the hidden-variable model), the degree sequence and the block structure or the expected degree sequence and the
block structure. These network ensembles can be considered as a generalization of the random graphs to more complex
network topologies.
Similarly, most of the generative models for multilayer networks can be divided also into two classes:
• growing multilayer networks models, in which the number of the nodes grows, and there is a generalized preferential
attachment rule [55,57,121–123]. These models explain multilayer network evolution starting from simple, and
fundamental rules for their dynamics.
• multilayer network ensembles, which are ensembles of networks with N nodes in each layer satisfying a certain set of
structural constraints [49,69,113,124]. These ensembles are able to generate multilayer networks with fully controlled
set of degree–degree correlations and of overlap.
In the case of multilayer networks the modeling framework is still in its infancy. Yet, different ways to construct multiplex
networks with different topological features have been proposed, and summarizing the main results and concepts coming
from these works is the object of the present section.
3.1. Multiplex networks models
3.1.1. Growing multiplex networks models

Growing multiplex networks’ models start from the consideration that many real networks are the result of a growing
process, such as, for instance, the case of the airport network formed by airline connections of different airline companies
[78,117], or that of the social network between scientists collaborating with each other and citing each other [113]. Similarly
to what happens for single-layer networks, the way the links of the new nodes are attached to the existing nodes is not
random, therefore it is important to explore the consequences on the resulting topology of assuming different attachment
kernels.
In two recent papers [55,57], a similar growing multiplex model has been proposed. There, the network has a dynamics
dictated by growth, and generalized preferential attachment. Starting at time t = 0 from a duplex network with n0 nodes
(with a replica in each of the two layers) connected by m0 > m links in each layer, the model proceeds as follows:
• Growth: At each time t ≥ 1 a node with a replica node in each of the two layers is added to the multiplex. Each newly
added replica node is connected to the other nodes of the same layer by m links.
• Generalized preferential attachment: The new link in layers α = 1, 2 is attached to node i with probability Πiα proportional
[1] [2]
to a linear combination of the degree ki of node i in layer 1 and ki of node i in layer 2, i.e.,
Πi [1]
∝ c1,1 ki + c1,2 ki ,
[1] [2]
Πi[2] ∝ c2,1 k[i 1] + c2,2 k[i 2] , (85)

where cα,β ∈ [0, 1] with c1,1 + c1,2 = c2,1 + c2,2 = 1.
In the case in which c1,1 = c2,2 = 1 the model reduces to two apparently decoupled BA models. Nevertheless, in
the multiplex network two replica nodes have the same age, therefore this characteristic of the model is responsible for
generating degree–degree correlations in the layers of the duplex. The model for c1,1 = c2,2 = 1 has been solved by the use
of the master equation approach [57] giving the joined degree distribution P (k, q) of having a node with degree k[1] = k in
layer 1 and degree k[2] = q in layer 2,
2Γ (2 + 2m)Γ (k)Γ (q)Γ (k + q − 2m + 1)

P (k, q) = . (86)
Γ (m)Γ (m)Γ (k + q + 3)Γ (k − m + 1)Γ (q − m + 1)
The average degree k̄(q) defined in Section 2.4.1 in one layer conditioned to the degree in the other layer [57] is given by
 m(q + 2)
k̄(q) = k[1] P (k[1] |k[2] = q) = . (87)
1+m
k[1]
Therefore, the two layer have positive degree correlations.

In this case, the degree distribution in each layer P (k) is given by the degree distribution of the BA network model.
In Ref. [57] it has been shown by mean-field calculations and simulations (see Fig. 13) that also for other values of the
parameters {cα,β } the average degree k̄(q) in one layer conditioned to the degree of the other layer remains linear, because
older nodes of the multiplex are more connected than younger nodes in both layers. Moreover, in Ref. [55] it has been
shown that for c1,1 < 1 and c2,2 < 1 the degree correlations in the two layers become more significant. In fact, the Pearson
correlation coefficient r defined in Section 2.4.1 between the degrees in the two layers has been characterized for the model
with c1,1 = c2,2 = 1 − ϵ and c1,2 = c2,1 = ϵ , finding in the limit t → ∞

1
⟨(k − ⟨k⟩)(q − ⟨q⟩)⟩  for ϵ = 0
r = = 2 (88)
σk σq 1 for ϵ > 0,
where k = k[1] indicates the degree in layer 1 and q = k[2] indicates the degree in layer two. Finally, both Refs. [55,57]
explore some variations of this model. In Ref. [57] the case of a semilinear attachment kernel (giving rise to networks with
different degree distributions) and that of time delay of the arrival of one replica node in one layer have been discussed. In
Ref. [55] an initial attractiveness has been added to the generalized preferential attachment, giving rise to two layers formed
by scale-free networks with a tunable value of the power-law exponent γ > 2.
A similar growing multiplex model has been also proposed by Magnani and Rossi in Refs. [122,123]. In their model, the
multiplex layers grow by the addition of new links and new nodes that do not arrive in the different layers at the same time.
Also in their case, the way the new links are attached to the rest of the network follows the preferential attachment rule.
The model displays layers that have either uncorrelated or positively correlated degree features.
In Ref. [121] it has been shown that growing multiplex networks models can also generate negative degree correlations
between the layers. In particular, the attachment kernel used in Ref. [121] is a non-linear kernel. The model includes growth
and non-linear preferential attachment depending on the degree of the same node in the different layers. For a duplex, the
model is described by the following algorithm. Starting at time t = 0 from a duplex network with n0 nodes (with a replica
a b
Fig. 13. (Color online) Panels (a)–(b): Degree distribution P (k) in each layer (left), and the interlayer degree–degree correlations k̄(k) (right) for the case
c1,1 = 1, c1,2 = 0, c2,1 = 0, c2,2 = 1 (top) and the case c1,1 = 0, c1,2 = 1, c2,1 = 1, c2,2 = 0 (bottom) attachment kernels (red circles are for the first layer,
yellow squares for the second layer).
in each of the two layers) connected by m0 > m links in each layer, the model proceed as follows.
• Growth: At each time t ≥ 1, a node with a replica node in each of the two layers is added to the multiplex. Each newly
added replica node is connected to the other nodes of the same layer by m links.
• Generalized non-linear preferential attachment: The new links are attached to node i with probability Πi[1] in layer 1 and
[2]
with probability Πi in layer 2 with
[β]
i qi ,
Πi[1] ∝ k[α]
[β]
Πi[2] ∝ q[α]
i ki (89)
where ki is the degree of node i in layer 1, and qi is the degree of node i in layer 2. Notice that here α and β are two
parameters of the model taking real values.
This model displays a very interesting physics, generating multiplex networks with different types of degree distributions,
degree–degree correlations, and with a condensation of the links for some parameter values. In Ref. [121], the phase diagram
of the model (see Fig. 14) has been completely characterized. The number of distinct degree classes |k| shown in panel (a)
of Fig. 14 allows to define a region of parameter space corresponding to homogeneous degree distributions (region Ia ), and
another region corresponding to heterogeneous degree distributions (region Ib ). The inverse partition ratio Y2−1 of the degree
distribution in a given layer shown in panel (b) of Fig. 14 provides evidence of a condensation occurring for α + β > 1 when
β ≥ 0, or for α > 1 when β < 0, in agreement with the theoretical derivations discussed in Ref. [121]. Finally, the Kendall
τ correlation coefficient between the degrees of corresponding replica nodes shows that the model can display at the same
time positive and negative degree–degree correlations in the two layers. Panel (d) of Fig. 14 shows the Kendall correlation
coefficient τ (t ) between the degree of corresponding replica nodes, calculated only over the nodes arrived at time t ′ < t.
This last panel shows that for every β < 0 the older nodes have always negatively correlated degrees in the two different
layers. Therefore, in the region in which τ > 0 and β < 0 shown in panel (c) of Fig. 14, only the recently added nodes in the
multiplex have positively correlated degrees, due to the stochastic nature in the model.
Other growing multiplex network models have been proposed [36,117], where the multiplex network grows by the
addition of an entire new layer at each time step. In Ref. [36], two nodes i and j in the new layer are linked with a probability
pij that depends on the quantity there called node multiplexity Qij . In particular, pij can be either positively correlated with
a4 120 b4 10000
9000 4c 1 d
3 100 3 8000 3
2 2 2 0.5
80 7000
1 1 6000 1
β 0 60 β 0 5000 β 0 0
-1 -1 4000 -1
40 3000
-2 -2 -2 -0.5
-3 20 -3 2000 -3
-4 -4 1000 -4
0 0 -1
-4 -3 -2 -1 0 1 2 3 4 -4 -3 -2 -1 0 1 2 3 4 -4 -3 -2 -1 0 1 2 3 4
α α α
Fig. 14. (Color online) Phase diagrams of the non-linear multiplex network growing model. By means of a color code, the following quantities are plotted
(see Ref. [121] for the definition of the plotted measures): (a) the number of distinct degree classes |k|, (b) the participation ratio Y2−1 , and (c) the Kendall’s τ
correlation coefficient. The solid black lines in panel (a) and (b) separate the non-condensed (region I) from the condensed phase (region II, small |k|, small
Y2−1 ). In region I we can have either homogeneous (region Ia ) or heterogeneous degree distributions (region Ib ). The solid black line in panel (c) separates
the two regions corresponding to positive (region +, τ > 0) or negative interlayer degree correlations (region −, τ < 0). The value of τ for the whole
multiplex is negative only in region −b . In panel (d) the plot of τ (t ), which is the Kendall’s restricted to the nodes arrived up to time t is shown. The dashed
black line corresponds to τ = 0 and is reported for visual reference. For β < 0 the interlayer correlations for older nodes are disassortative (τ (t ) < 0 for
small t), even if the value computed on all the nodes might be positive due to the prevalence of young nodes.
Qij , or negatively correlated to it. In the first case two nodes that are active at the same time in many layers are more likely
to be connected in the new layer, in the second case two nodes that are active at the same time in many layers have small
probability to be connected in the new layer. In Ref. [117], instead, every node i of the new layer will be active with a
probability Pi proportional to the activity of the node Bi , i.e.,
Pi ∝ A + Bi (t ), (90)
where A is a parameter of the model. This model enforces a sort of ‘‘preferential attachment’’ of the new layers to nodes of
high activity Bi , and a power-law distribution P (Bi ) of the activities of the nodes.
3.1.2. Simple static models

The simplest way to obtain a static generative model for multiplex networks is to generalize the existing methods
for single-layer ones. Fixing the degree sequence in each layer, one can use a configuration model to obtain a particular
realization of the given set of connectivities. Of course, the full structural information in a multiplex includes also the
connections between layers. In Refs. [48,125], the Authors have made the choice to add interlayer links arbitrarily. A different
approach is to keep using a configuration model, but to specify the edges between the layers by means of a joint-degree
distribution [56,126,127]. Note that, while the terms ‘‘joint-degree sequence’’, ‘‘joint-degree matrix’’ and ‘‘joint-degree
distribution’’ are normally used in literature to indicate quantities related to the degrees of the nodes at the end of the
links in a simple network, here they explicitly refer, instead, to multilayer connections [128].
A similar method is to specify the degree sequences together with a probability matrix whose element (i, j) is the fraction
of interlayer links between layers i and j. The actual link placement is still achieved via uniform random choice [129–133].
A generalization of this approach has been proposed in Ref. [134]. There, the Authors impose the degree correlations within
and between layers by means of a set of matrices that specify the fraction of edges between nodes of given degrees in given
layers.
Further models of multilayer networks can be created by defining classes of nodes that share similar structural
characteristics. While the definitions of such classes are somewhat arbitrary, these models find application in the detection
of higher-level features in real-world networks, mostly related to community structure [135–138].
The choice of any network model for a particular study results in the imposition of constraints of some sort. Each network
model corresponds to a network ensemble, i.e., to the set of networks that satisfy the constraints implicitly imposed by the
model. Thus, it is important to study the properties of network ensembles, which we discuss in the following from a statistical
physics perspective.
3.1.3. Ensembles of multiplex networks

Network ensembles and their entropy. A given set of constraints can give rise to a microcanonical network ensemble, satisfying
the hard constraints, or to a canonical network ensemble in which the constraints are satisfied in average over the ensemble.
The canonical network ensembles have been widely studied in the sociological literature under the name of exponential
random graphs or p⋆ models [139–142], and recently studied by physicists [143–147]. The most popular examples of
microcanonical network ensembles are the G(N , L) ensemble and the configuration model. The distinction between the
canonical and microcanonical ensembles allows to draw a well defined parallelism [146] between such network ensembles
and the classical ensembles in statistical mechanics where one considers system configurations compatible either with a
fixed value of the energy (microcanonical ensembles) or with a fixed average of the energy determined by the thermal bath
(canonical ensembles). For example, the G(N , L) random graphs formed by networks of N nodes and L links is an example
of a microcanonical network ensemble, while the G(N , p) ensemble (where each pair of links is connected with probability
p) is an example of canonical network ensemble since the number of links can fluctuate but has a fixed average given by
⟨L⟩ = pN (N − 1)/2. Despite this similarity, an ensemble that enforces an extensive number of hard constraints on its
networks (such as the degree sequence) is not equivalent in the thermodynamic limit to the conjugated network ensemble
enforcing the same constraints in average (such as the sequence of expected degrees). For instance, the ensemble of regular
networks with degrees ki = c ∀i = 1, 2, . . . , N is clearly not equivalent to the Poisson network in which every node has the
same expected degree k̄i = c. The non-equivalence of microcanonical and canonical network ensembles with an extensive
number of constraints needs also to be taken into account when characterizing dynamical processes on networks, because
it can have relevant effects on their critical behavior.
The entropy of network ensembles is given by the logarithm of the number of typical networks in the ensemble, and it
quantifies the complexity of the ensemble. In particular we have that the smaller is the entropy of the ensemble, the smaller
is the number of networks satisfying the corresponding constraints, implying that these networks are more optimized. Both
the network ensembles and their entropy can be used on several inference problems to extract information from a given
network [147,148].
Multiplex network ensembles. The statistical mechanics of randomized network ensembles (exponential random graphs)
has been extended to describe multiplex ensembles [49,69,113,124,149], and applied to analyze different multiplex data
sets [113,150,151]. Consider a multiplex formed by N labeled nodes i = 1, 2, . . . , N and M layers. We can represent the
multiplex as described for example in Ref. [68]. To this end, we indicate by G⃗ = (G1 , G2 , . . . , GM ) the set of all the networks
Gα at layer α = 1, 2, . . . , M forming the multiplex. Each of these networks has an adjacency matrix with matrix elements
aαij = 1 if there is a link between node i and node j in layer α = 1, 2, . . . , M, and zero otherwise. A multiplex ensemble
is specified when the probability P (G ⃗ ) for each possible multiplex is given. In a multiplex ensemble, if the probability of a
multiplex is given by P (G
⃗ ), the entropy of the multiplex S is defined as

S=− P (G
⃗ ) log P (G
⃗) (91)
⃗
G
and measures the logarithm of the typical number of multiplexes in the ensemble. One can distinguish between
microcanonical and canonical multiplex network ensembles, and between uncorrelated and correlated multiplex network
ensembles [49].
Uncorrelated multiplex networks ensembles. The uncorrelated multiplex ensembles are ensembles of networks in which
the links in different layers are not correlated. Note that the degree in the different layers can nevertheless be eventually
correlated, as described below. The probability P (G⃗ ) of an uncorrelated multiplex ensemble is given by
M
Pα (Gα ),

P (G
⃗) = (92)
α=1
where Pα (Gα ) is the probability of network Gα on layer α . In this case, the entropy of the multiplex ensemble is the sum of
the entropies of the network in the single layers, i.e.,
M M
Sα = − Pα (Gα ) log Pα (Gα ).
 
S= (93)
α=1 α=1
In a uncorrelated multiplex network, the links in two different layers α and α ′ ̸= α are uncorrelated. In fact, from Eq. (92)
it immediately follows that
⟨aαij aαij ⟩ = ⟨aαij ⟩⟨aαij ⟩

′ ′
(94)
for every choice of pair of nodes i, j.
Examples of uncorrelated multiplex ensembles are multiplex layers with given degree sequence, or given degree
sequence and community structure, directed networks with given in–out degree, weighted networks with given degree
and strength of the links, etc.
A pivotal role is played by the ensemble in which we fix either the expected or the exact degree sequence. These degree
sequences can be eventually correlated. Let us consider here a simple case of a duplex M = 2 in which we either fix (i)
the expected degree sequence (canonical multiplex network ensemble) or (ii) the exact degree sequence (microcanonical
multiplex network ensemble).
• (i) Fixed expected degree sequence
In the first case, when we fix the expected degree sequence {k̄αi }, the probability of the multiplex PC (G
⃗ ) is given by
M 
pαij aαij + (1 − pαij )(1 − aαij )

PC (G
⃗) =
 
(95)
α=1 i<j
where pαij is the probability of the link (i, j) in layer α and can be expressed in terms of the Lagrangian multipliers λαi as
−(λαi +λαj )
α e
pij =
−(λαi +λαj )
. (96)
1+e
The Lagrangian multipliers λαi are fixed by the NM conditions
k̄αi = pαij ,

(97)
j̸=i
where k̄αi is the expected degree of node i in layer α . This last expression, in the case in which there is a structural cutoff
k̄αi <

⟨k̄α ⟩N, has a very simple solution given by
k̄αi k̄αj
pαij = . (98)
⟨k̄α ⟩N
Note that here ⟨ ⟩ indicates average over the nodes. This is an example of canonical multiplex networks. The entropy of
this multiplex ensemble is called Shannon entropy and is given by

S = − PC (G
⃗ ) ln PC (G
⃗)
⃗
G
M
[pαij log pαij + (1 − pαij ) log(1 − pαij )].

=− (99)
α=1 i<j
These ensembles can display degree correlations in the different layers, but they are uncorrelated ensembles in the sense
that they satisfy Eqs. (92) and (94).
The multiplex networks in this ensemble can be generated by putting a link between node i and node j with probability
pαij given by Eq. (96) or, in presence of the structural cutoff, by Eq. (98).
• (ii) Fixed degree sequence
Here we also consider the microcanonical multiplex network ensembles in which we fix the exact degree sequence
{kαi } in every layer. In this case the probability of a multiplex network is given by
 
M
1  α

α
PM (G
⃗) = δ ki , aij , (100)
Z α=1 i<j j
where Z is a normalization constant. We call the entropy of this ensemble Gibbs entropy N Σ to distinguish it from the
Shannon entropy S of the canonical ensembles. We have

NΣ = PM ( G
⃗ ) ln PM (G
⃗ ),
⃗
G
= S − NΩ (101)
where S is the Shannon entropy of the conjugated canonical model given by Eq. (99) calculated assuming that the
expected degree of each node i is given exactly by its degree ki , and where Ω is given by
M 
N
log πkαi (kαi )

NΩ = − (102)
α=1 i=1
where πy (x) = x1! yx exp[−y] is the Poisson distribution with average y calculated at x. Therefore, the Gibbs entropy
N Σ is lower than S, and the microcanonical ensemble is not equivalent in the thermodynamic limit to the conjugated
canonical ensemble. The multiplex networks in this ensemble can be generated by generating the different networks in
the different layers with the configuration model without allowing for multiple edges and tadpoles in the networks.
These types of networks are the most simple examples of multiplex, and many dynamical processes have been first studied
on these ensembles, by analyzing also the effect of correlation between the degrees in different layers. In Ref. [49] it has been
shown that if a uncorrelated multiplex network ensemble is formed by sparse networks (e.g. Poisson networks, exponential
networks with finite average degree, scale-free networks with power-law exponent γ > 2), the expected overlap of the
links between any two layers of the multiplex is negligible. Therefore, these ensembles cannot be used to describe multiplex
networks where the overlap of the links is significant.
Correlated multiplex networks ensembles. The correlated multiplex ensembles are ensembles of networks in which the links
in different layers are correlated. In particular, in the correlated multiplex network ensembles, Eq. (92) does not hold, i.e.
M
Pα (Gα ).

P (G
⃗ ) ̸= (103)
α=1
In this correlated ensembles, the entropy S cannot be expressed as the sum of the entropies S α , i.e.
M
Sα = − Pα (Gα ) log Pα (Gα ).
 
S ̸= (104)
α α=1
Furthermore, one has that it exists at least a pair of layers α and α ′ and a pair of nodes i and j for which
⟨aαij aαij ⟩ ̸= ⟨aαij ⟩⟨aαij ⟩.

′ ′
(105)
Examples of correlated multiplex ensembles are multiplex layers with given multidegree sequence, directed networks
with given directed multidegree sequence, weighted networks with given multidegree and multistrength of the links,
etc. [49,113].
In particular, correlated multiplex networks can display a significant overlap of the links. Let us consider here a simple
case of a duplex M = 2 in which we either fix (i) the expected multidegree sequence (canonical multiplex network
ensemble) or (ii) the exact multidegree sequence (microcanonical multiplex network ensemble).
• (i) Fixed expected multidegree sequence
The probability PC (G
⃗ ) of a multiplex network G
⃗ in which we fix the expected multidegree sequence {k̄m⃗
i } is given by
 
PC (G
⃗) = pm ⃗ m ⃗
ij Aij (106)
i <j ⃗
m
where pm⃗
ij ⃗ = (m1 , m2 , . . . , mα ) between node i and node j, and Am⃗ is the
is the probability of the multilink m
aαij mα + (1 − aαij )(1 − mα ) . For a sparse network, the probabilities
⃗
M  
multiadjacency matrix with elements Am
ij = α=1
pm⃗
⃗ ̸= 0⃗ can be expressed in terms of the Lagrangian multipliers λm
ij of a multilink m
⃗
i as
−(λm⃗ +λm⃗)
e i j
pm⃗
ij =  ⃗ +λm⃗)
, (107)
−(λm
e i j
⃗
m
⃗
ij = 1. The Lagrangian multipliers λi are fixed by the conditions
0 ⃗ ⃗
pm m

while pij is fixed by the normalization condition ⃗
m

ij ,
⃗ ⃗
k̄m
i = pm (108)
j̸=i
where k̄m⃗
⃗ ̸= 0⃗ of node i. This last expression, in the case in which there is a structural
i is the expected multidegree m 
i <
⃗
cutoff on the multidegree k̄m ⟨k̄m⃗ ⟩N has a very simple solution given by
⃗ m ⃗
k̄m
i k̄j
pm⃗
ij = , (109)
⟨k̄m⃗ ⟩N
⃗ This is an example of canonical multiplex networks. The entropy of this multiplex ensemble
⃗ ̸= 0.
for every multilink m
is called again Shannon entropy and is given by

S = − PC (G
⃗ ) ln PC (G
⃗)
⃗
G

⃗ ⃗
=− [p m m
ij log pij ]. (110)
⃗
m i<j
In order to construct a correlated multiplex ensemble with given multidegree sequence it is sufficient to follow the
following scheme.
⃗
– Calculate the probability pm ⃗ between node i and j using Eq. (107) or directly Eq. (109) in the
ij to have a multilink m
presence of a structural cutoff on the multidegrees.
ij and consequently put a link in every layer α
⃗
– For every pair of node i and j, draw a multilink m ⃗ with probability pm
where mα = 1, and put no link in every layer α where mα = 0.
• (ii) Fixed multidegree sequence

The conjugated multiplex network ensemble of the previous example is the ensemble in which we fix the exact
multidegree sequence {kαi } in every layer. In this case the probability of a multiplex network is given by
 
1  
PM (G) =
⃗ δ ki ,
m⃗
aij ,
m⃗
(111)
Z ⃗ i <j j
m
with Z being a normalization constant. We call again the entropy of this ensemble Gibbs entropy N Σ to distinguish it
from the Shannon entropy S of the canonical ensemble. We have

NΣ = PM (G
⃗ ) ln PM (G
⃗ ),
⃗
G
= S − NΩ (112)
where S is the Shannon entropy of Eq. (99) calculated assuming that the expected multidegrees of a node are given by
the exact multidegree of the nodes in the microcanonical ensemble, and where Ω is given by
 N

NΩ = − log πkm⃗ (km
i ).
⃗
(113)
i
M i =1
mα >0

⃗|
m
α=1
Spatial multiplex network ensembles. When the multiplex networks are embedded in a real or in a hidden space (as in the case
of the airport multiplex network in Ref. [78]), the overlap of the links can emerge naturally from the correlations induced
by the distance in the embedding space. Intuitively, if in every layer links are more likely between nodes that are closer
in the embedding space, then one can observe non-negligible overlap of the links because two nodes that are close in the
embedding space are more likely to be connected to each other in every layer of the multiplex. We have that, if the probability
P (G
⃗ |{⃗ri }) of a network G
⃗ with N nodes on positions {⃗ri } is given by
M
Pα (Gα |{⃗

P (G
⃗ |{⃗ri }) = ri }), (114)
α=1
where P (Gα |{⃗

ri }) is the probability of a network in layer α , then we can observe a significant overlap of the links [149].
3.2. Ensembles of networks of networks
In this case, there are multiple degrees of freedom that may characterize these complex structures, and therefore several
types of networks of networks can be considered:
• Networks of networks with fixed supernetwork, i.e., networks of networks where if a layer α is connected with layer β ,
every node (i, α) is connected to its replica node (i, β) in layer β [152,153].
• Networks of networks with given superdegree distribution, where every node (i, α) is connected to q = qα replica nodes
(i, β) of randomly chosen layers β [154].
• Networks of networks with fixed supernetwork and random permutations of the labels of the nodes, i.e. network of networks
where if a layer α is connected with layer β , any node (i, α) is connected to a single node (j, β), where j is taken from a
random permutation πα,β of the indices {i} [26,155,156].
α,β
• Networks of networks with multiple interconnections, where every node (i, α) has ki connections with nodes in layer
β [80].
In order to distinguish the links between nodes of the same layers and the links between nodes of different layers, we will
call the latter interlinks.
3.2.1. Case I: networks of networks with fixed supernetwork and links allowed only between replica nodes
In this subsection we will consider a network of networks in which if network α is connected with network β , each node
(i, α) of network α is connected to the node (i, β) of network β (see Fig. 15). The supra-adjacency matrix of the network of
networks has elements aiα,jβ = 1 if there is a link between node (i, α) and node (j, β), and zero otherwise. In these specific
networks, aiα,jβ = 0 if both i ̸= j and α ̸= β . The network of networks is characterized by an adjacency matrix Aαβ , such
that for every node i = 1, 2, . . . , N, Aαβ = aiα,iβ . The adjacency matrix Aαβ is the adjacency matrix of the supernetwork, G,
of the network of networks. We can consider ensemble of networks where the interlinks between different layers are fixed
by the adjacency matrix Aαβ , but the network in a given layer is random. In particular, we can choose a canonical ensemble
Fig. 15. (Color online) Schematic view of a typical network of networks with given supernetwork. Interdependencies (interlinks between nodes from
different levels) are shown by the black dashed lines. Intralinks between nodes within layers are shown as solid red lines.
of networks of networks in which each layer is a uncorrelated network and each node has a given expected degree k̄αi in
each layer α . In this case, the probability P ({a}) of the supra-adjacency matrix a is given by
 
 N
M  
P ({a}) = δ(aiα,iβ , Aαβ ) δ(aiα,jβ , 0) (115)
α=1 i=1 β̸=α j̸=i
k̄αi k̄αj k̄αi k̄αj

    
M 

× aiα,jα + 1− (1 − aiα,jα ) . (116)
α=1 i<j ⟨k̄α ⟩N ⟨k̄α ⟩N
An alternative is to consider an ensemble of networks of networks in which the network in each layer is built using the
configuration model, i.e. each node has given degree kαi in layer α . In this case, the probability of the supra-adjacency matrix
a is
   
M N M 
N N
1   
α

P ({a}) = δ(aiα,iβ , Aαβ ) δ(aiα,jβ , 0) δ ki , aiα,jα . (117)
Z α=1 i=1 β̸=α j̸=i α=1 i=1 j =1
3.2.2. Case II: networks of networks with given superdegree distribution

In this case, each node (i, α) can be linked only to its replica nodes (i, β), but there is no fixed supernetwork between
the layers. Nevertheless, we assume that each layer α has a given superdegree qα , i.e. in these networks of networks every
node (i, α) is linked to q = qα replica nodes (i, β) of randomly chosen layers β [154] (see Fig. 16).
Moreover, each network (layer) α is generated randomly according to a canonical or a microcanonical network ensemble.
If fact, we can consider a microcanonical version of this model in which the degree sequence in each layer {kαi } is fixed, or
a canonical version of this model in which the expected degree sequence {k̄αi } is fixed. Therefore, for the canonical network
of networks ensemble, every network of networks with a supra-adjacency matrix a has a probability P (a) given by

        
M N
1    
P (a) = δ qα , (1 − aiα,jα ) δ aiα,jβ , 0 .

aiα,iβ aiα,jα + 1− (118)
Z α=1 i=1 β̸=α ⟨k̄α ⟩N ⟨k̄α ⟩N β̸=α j̸=i
Alternatively, one can consider the microcanonical network of networks ensemble in which every network of networks with
a supra-adjacency matrix a has a probability P (a) given by
    
M N M 
N 
1  α
  
P (a) = δ ki , δ qα , δ aiα,jβ , 0 .
 
aiα,jα aiα,iβ (119)
Z α=1 i=1 j̸=i β̸=α α=1 i=1 β̸=α j̸=i
Fig. 16. (Color online) Schematic representation of a network of networks with given superdegree distribution. Each layer α has assigned given superdegree
qα but each node (i, α) can be linked to any qα replica nodes.
3.2.3. Case III: networks of networks with fixed supernetwork and random permutations of the labels of the nodes
We here consider networks of networks in which if network α is connected with network β , each node (i, α) of network
α is connected to a single node (j, β) of network β randomly chosen between the nodes of layer β (see Fig. 17). The one
to one mapping between the nodes is performed by defining a permutation πα,β of the indices {i} such that j = πα,β (i)
if and only if node (i, α) is linked to node (j, β). In order to define a undirected network of networks, we need to impose
that the permutation πβ,α is the inverse permutation of πα,β , enforcing that if j = πα,β (i) then i = πβ,α (j). The network of
networks supra-adjacency matrix has elements aiα,jβ = 1 if there is a link between node (i, α) and node (j, β), i.e. if and
only if j = πα,β (i), and zero otherwise. In these specific networks, for β ̸= α , aiα,jβ = 0 if j ̸= πα,β (i). The network of
networks is characterized by an adjacency matrix Aαβ , such that for every node i = 1, 2, . . . , N, Aαβ = aiα,πα,β (i)β is the
adjacency matrix of the supernetwork.
The present case differs substantially from case I. For instance, consider the case of a supernetwork forming a loop of
size M between the different layers. In case I, starting from each node (i, α) and following only interlinks between different
layers, we can reach only M other nodes, while in case III (as each permutation πα,β (i) is random) the number of different
nodes that can be reached following interlinks can be significantly higher than M (see Fig. 17). If the supernetwork is a fully
connected network of M nodes, case I describes a multiplex with all the replica nodes in the different layers connected to
each other, while case III does not reduce to the multiplex case.
One can consider an ensemble of networks where the interlinks between different layers are fixed by the adjacency
matrix Aαβ and the permutations πα,β (i), but the network in a given layer is random. In particular, we can choose a canonical
ensemble of networks of networks in which each layer is a uncorrelated network, and each node has a given expected degree
k̄αi in each layer α . In this case, the probability P ({a}) of the supra-adjacency matrix a is given by
 
M 
 N    
P ({a}) = δ(aiα,πα,β (i)β , Aαβ ) δ(aiα,jβ , 0) (120)
α=1 i=1 β̸=α j̸=πα,β (i)
 

    
M 

× aiα,jα + 1− (1 − aiα,jα ) . (121)
α=1 i<j ⟨k̄α ⟩N ⟨k̄α ⟩N
An alternative is to consider an ensemble of networks of networks in which the network in each layer is built using the
configuration model, i.e. each node has given degree kαi in layer α . In this case, the probability of the supra-adjacency matrix
Fig. 17. (Color online) Schematic representation of a network of networks with fixed supernetwork and random permutations of the labels of the nodes
(case III). The supernetwork can be arbitrary but in the drawing we show a specific realization in which the supernetwork is a loop formed by M = 3 layers.
a is
   
M N  M  N N
1 

α
  
P ({a}) = δ(aiα,πα,β (i)β , Aαβ ) δ(aiα,jβ , 0) δ ki , aiα,jα . (122)
Z α=1 i=1  β̸=α j̸=πα,β (i)
 α=1 i=1 j =1
3.2.4. Case IV: networks of networks with multiple interconnections

The last example assumes that the nodes of a layer can be linked to any other node in other layers, and that also multiple
interlinks between one node and the node of other layers are allowed. This model has been used, for instance, in Ref. [80].
αβ
The model considers every node (i, α) as having ki connections with nodes in layer β = 1, 2, . . . , M, and therefore it
α,β
requires a number MN of parameters (the degrees ki ) which is much larger than the number of parameters needed for
the other models proposed before, as long as M ≫ 1. The typical structure of a network of networks in this ensemble is
αβ
shown in Fig. 18. Once again, one can have a microcanonical version of this model (in which the degree sequences {ki }
αβ
are fixed), or a canonical version (in which the expected degree sequences {k̄i } are fixed). For the canonical ensemble the
supra-adjacency matrix a has a probability P (a) given by
αβ β,α αβ β,α
    
M 
N 
 k̄i k̄j k̄i k̄j
P (a) = aiα,jβ + 1− (1 − aiα,jβ ) , (123)
α=1 β≤α i=1 j<i ⟨k̄αβ ⟩N ⟨k̄αβ ⟩N
while for the microcanonical ensemble the probability P (a) is
  
M M N
1  αβ

P (a) = δ ki , aiα,jβ . (124)
Z α=1 β=1 i=1 j̸=i
4. Resilience and percolation
The present section is devoted to review the rather huge literature regarding two very relevant, and somehow correlated,
processes: resilience of the multilayer networks’ structure under random failures (and/or cascades of failures), and
percolation.
Resilience is the ability of a network to maintain its main topological structure under the action of some kind of damages.
It is well known that single-layer heterogeneous networks, such as scale-free (SF) networks, are particularly resilient to
random damages, whereas they generally are extremely fragile to attacks targeting the nodes with higher degrees. This is,
actually, one important motivation that the scientific community has given to the observed SF degree distribution: both
technological (e.g. the Internet) and biological networks (e.g. the protein–protein interaction and other biological networks
in the cell) display a SF nature because they must be robust against random failures [157–163]. In a multilayer scenario,
nevertheless, conclusions and predictions are far different, and indeed the available literature allows one to conclude that,
if a network is interdependent with other networks, its robustness properties can be strongly affected.
On the other side, percolation (from the Latin percolare, whose meaning is ‘‘to filter’’ or ‘‘to trickle through’’) refers, in
classical physics, to the movement and filtering of fluids through porous materials. Within the last fifteen years, the main
αβ
Fig. 18. (Color online) Schematic representation of a network of networks with multiple interconnections. Each node (i, α) has ki links with nodes in
layer β .
formalism and predictions of percolation theory have been applied also to complex networks, for the study of the emergence
(and robustness) of a giant connected component in the graph having a size of the order of the size of the entire network.
In this framework, statistical physics concepts (such as scaling theory and phase transitions) have been used to characterize
network percolation properties, and study percolation thresholds. Along this section we will extensively discuss how the
traditional percolation theory has been recently generalized to the context of multilayer networks.
4.1. The fragility of multilayer interdependent networks
It is a matter of fact that in many multilayer networks some nodes of a layer are interdependent on nodes in other
layers. A node is interdependent on another node in another layer if it needs the other node to function in order to
function itself properly. Examples of interdependent networks are found in complex infrastructures, which are increasingly
interconnected, in economy where networks of banks, insurance companies, and firms interact with each other, in
transportation networks, where railway networks, airline networks and other transportation networks are interdependent,
or in biology where the cell is alive only if all its biological networks are functioning at the same time.
When two or more networks are interdependent, a fraction of node failures in one layer can trigger a cascade of failures
that propagate in the multilayer network (as it was the case, for instance of the 2003 blackout in Italy [25]), and this is
indicative of an increasing fragility of these structures with respect to the single-layer cases [25,164]. Therefore, new aspects
of the robustness of networks emerge when interdependencies are taken into account. Moreover, the impact of the topology
of the interdependent networks on their robustness properties is also unexpected: in fact, in absence of correlations,
interdependent SF networks are more fragile than Poisson networks when a constant average degree is kept [25]. When
we consider multiplex networks or networks of networks with given supernetwork and links allowed only between replica
nodes as described in Section 3.2.1, the system responds globally to the damage inflicted in a single layer. At any given time,
either all the layers have a percolating cluster, or none of them has one [152]. When instead one considers a network of
networks with given superdegree sequence as described in Section 3.2.2, layers with different number of interdependencies
respond differently to the inflicted damage. In this latter case, the layers with higher number of interdependencies, i.e. with
higher superdegree qα , are more fragile than those with smaller superdegree [154]. Indeed, in these layers the percolation
cluster disappears for a fraction of initial damaged nodes 1 − p smaller than the one that is necessary to disrupt layers with
smaller number of interdependencies. Finally, if we consider a network of networks with fixed supernetwork and random
permutations of the label of the nodes, as described in Section 3.2.3 one has to distinguish between the case in which the
supernetwork is a tree, and that in which the supernetwork contains loops [26,156]. If the supernetwork is a tree, all the
layers start to percolate when the fraction of initially damaged nodes is less than 1 − pc , i.e. either all the layers percolate or
none of them does [153]. In the case in which the supernetwork contains loops and the supernetwork is a random network
with given superdegree distribution, the layers α with higher superdegree qα are more fragile than those with smaller
superdegree. In this case the mutually connected component is described by the same equations derived for the network of
networks described in Section 3.2.2 [154].
The increased fragility of interdependent networks opens new important scientific questions. A major one is the design of
optimal strategies to mitigate the dramatic effect of interdependencies and allow for correlated topologies with increased
a b c d
Fig. 19. Schematic description on how to find the mutually connected component in a multilayer interdependent network (duplex) considering the
cascading failures propagating from one layer to the other one as described in Section 4.2. The set of active nodes that remains at the end of the cascade is
the mutually connected component.
Source: Reprinted figure from Ref. [25]. Courtesy of S. Havlin.
resilience under random damage. This might have consequence for the infrastructure design of the new generation, for
reducing risk in financial networks, and for assessing the robustness of the biological networks in the cell. Notice that, while
for random interdependent networks percolation is clearly sharpened by the presence of interdependencies, the question
weather interdependencies always sharpen the percolation transition has been matter of debate [165–167].
4.2. Percolation in interdependent networks
After the pioneering and seminal work of Ref. [25], it became clear that, in presence of interdependencies, the robustness
of multilayer networks can be evaluated by calculating the size of their mutually connected giant component (MCGC) when
a random damage affects a fraction 1 − p of the nodes in the system. The MCGC of a multilayer network is the largest
component that remains after the random damage propagates back and forth in the different layers.
In Fig. 19, it is shown how to construct the MCGC for the case of a multiplex formed by two layers (layer A and layer
B) where the replica nodes are interdependent. Assume that a group of nodes in layer A is damaged, all the nodes in the
giant component of layer A are active while the nodes that are not in the giant component are not active. Then consider
the layer B. All the nodes of layer B that are interdependent with nodes that are not active in layer A, are damaged. Next,
set as active all the nodes that remain in the giant component of layer B, and set as not active all the nodes that are not in
the giant component. By repeating the algorithm alternating the analysis of layer A and layer B, it is possible to characterize
an avalanche of failure events that propagates from one layer to the other one, until the avalanche does not propagate any
more. The set of active nodes that remains at the end of the iteration is the MCGC. Between any two nodes in this mutually
connected component there are at least two paths, one in one layer A and one in layer B, that connect the two nodes and
that pass only through nodes that belong to the mutually connected component. In this case, the MCGC is also called the
viable cluster of the network [168].
The size of the MCGC of a multilayer network emerges discontinuously at a critical value p = pc , at least in the case
of random networks in which there is no overlap of the links in the different layers and every node is interdependent at
least on another node in another layer. This is in contrast with the theory of percolation in single-layer networks, where
the giant component emerges continuously at a second order phase transition. Moreover, if one approaches the value
p = pc , the interdependent networks are affected by cascading failures that propagate throughout the structure. Other works
have extended the original formalism to network of networks [26,153,156], network with partial interdependencies [35]
with multiple dependencies links [169] and exploring the effect of targeted attack [170]. Subsequent studies on the
subject have been carried out by Son et al. [171]. The Authors call this algorithm an ‘‘epidemic spreading’’ process, but
the algorithm can be more suitably termed as the ‘‘message passing algorithm’’ for this generalized percolation problem
[172–174]. Here we will try to present all this material in a pedagogical way, starting from a generalization of the algorithm
of Ref. [171] that can be used both for multiplex and for networks of networks.
4.2.1. The mutually connected giant component of a multilayer network

In this subsection, we will consider a special type of a multilayer network formed by M layers α = 1, 2, . . . , M, each
one formed by interactions between N nodes i = 1, 2, . . . , N. Every node will be indicated by the pair (i, α), which denotes
the label i of the nodes in the specific layer α . We furthermore call all the nodes (i, α) characterized by the same label i
but belonging to different layers α the ‘‘replica nodes’’. Every node (i, α) can be connected to nodes (j, α) within the same
layer with ‘‘connectivity links’’, or with its ‘‘replica nodes’’ (i, β) in other layers with ‘‘interdependency links’’. In Ref. [169],
it is further allowed to one node in layer α to be interdependent to more than one node in a given layer β , but here (if not
explicitly otherwise stated) we make the simplifying assumption that every node in layer α can be interdependent at most
to a node in layer β , i.e. on its ‘‘replica node’’. This framework allows to treat at the same time, multiplex networks and
networks of networks as described in Sections 3.2.1 and 3.2.2. We define the network of networks with a supra-adjacency
matrix of elements aiα,jβ = 1 if there is a link between node (i, α) and node (j, β) and zero otherwise. Therefore, here we
will treat multilayer networks in which aiα,jβ = 0 if both i ̸= j and α ̸= β .
The MCGC is defined in Ref. [171] as the set of nodes (i, α) that satisfy the following recursive set of equations:
(a) at least one neighbor (j, α) of node (i, α) in layer α is in the MCGC;
(b) all the interdependent nodes (i, β) of node (i, α) are in the mutually connected giant component.
For a given multilayer network which is locally tree-like, it is easy to construct a ‘‘message passing’’ algorithm [172–174] that
allows us to determine if node (i, α) is in the mutually connected component. This approach has been proposed in Ref. [171]
and further investigated in Ref. [175]. We denote by σiα→jα = 1, 0 the message within a layer, from node (i, α) to node
(j, α), σiα→jα = 1 indicating that node (i, α) is in the mutually connected component when we consider the cavity graph
constructed by removing the link (i, j) in layer α . Furthermore, let us denote by Si′α→iβ = 0, 1 the message between the
‘‘replicas’’ (i, α) and (i, β) of node i in layers α and β . Si′α→iβ = 1 indicates that the node (i, α) is in the mutually connected
component when we consider the cavity graph by removing the link between node (i, α) and node (i, β). In addition, we
indicate with siα = 0 a node that is removed as an effect of the damage inflicted to the network, otherwise siα = 1.
If the node (i, α) is not damaged, i.e. siα = 1, the message σiα→jα will be equal to one if all the messages coming to node
(i, α) from the interdependent nodes (i, β) are equal to one, and if at least one message σℓα→iα coming from a node (ℓ, α)
neighbor of node (i, α) in layer α and different from node (j, α), is equal to one. If the node (i, α) is instead damaged, the
message σiα→jα will be equal to zero.
If the node (i, α) is not damaged, i.e. siα = 1, the message Si′α→iβ will be equal to one if all the messages coming to node
(i, α) from the interdependent nodes (i, γ ) different from (i, β) are equal to one, and if at least one message σℓα→iα coming
from a node (ℓ, α) neighbor of node (i, α) in layer α , is equal to one. If the node (i, α) is instead damaged, the message Si′α→iβ
will be equal to zero.
Therefore, the message passing equations are of the following form:
 
 
σiα→jα = siα Si′β→iα 1− (1 − σℓα→iα ) ,
β∈Ni (α) ℓ∈Nα (i)\j
 
 
Si′α→iβ = siα Si′γ →iα 1− (1 − σℓα→iα ) , (125)
γ ∈N (α)\β ℓ∈Nα (i)
where Nα (i) indicates the set of nodes (ℓ, α) which are neighbors of node i in layer α , and Ni (α) indicates the layers β of
the nodes (i, β) interdependent on the node (i, α).
Finally, Siα = 1 (Siα = 0) indicates that the node (i, α) is (is not) in the mutually connected component, namely
 
 
Siα = siα Si′β→iα 1− (1 − σℓα→iα ) . (126)
β∈Ni (α) ℓ∈Nα (i)
In other words, a node (i, α) is in the mutually connected component if it is not damaged (i.e. siα = 1), if all the messages
Si′β→iα coming from the interdependent nodes are equal to one, and if at least one message coming from a neighbor node
(ℓ, α) in layer α is equal to one. In Ref. [152], it has been shown that these message passing equations can be solved explicitly
for every network of networks of this type including both cases described in Sections 3.2.1 and 3.2.2 giving, for the messages
σiα→jα within each layer
    
σiα→jα = siα 1 − (1 − σℓα→iα ) (1 − σℓβ→iβ ) ,
  
siβ 1− (127)
ℓ∈Nα (i)\j β∈Ci (α)\α ℓ∈Nβ (i)
where Ci (α) is the set of layers β such that the ‘‘replica nodes’’ (i, β) are connected to the node (i, α) by paths going through
the interdependency links. Finally, the Eqs. (126) for Siα can be written exclusively as functions of the messages σiα→jα as
  
   
Siα = siα siβ 1 − (1 − σℓβ→iβ ) . (128)
β∈Ci (α) ℓ∈Nβ (i)
 
For the network of networks described in Section 3.2.3 or Section 3.2.4 a similar calculation can be carried out showing
that in networks of networks of any type, the mutually connected component can be defined as the following: A node (i, α)
is in the mutually connected giant component if and only if:
(a) it has at least one neighbor node (j, α) in layer α that belongs to the mutually connected component;
(b) all the nodes (ℓ, β) that can be reached from node (i, α) by interdependent links have at least one neighbor node that
belongs to the mutually connected component.
4.3. Interdependent multiplex networks
A multiplex network can be considered as a network of networks of the type described in Section 3.2.1, in which each
layer is connected with all the other layers (the supernetwork formed by the layer is a fully connected network) and each
node is connected only to its replica nodes in the other layers. Let us consider a multiplex network in which every layer α
has a degree distribution Pα (k) and in which every node (i, α) is interdependent on every one of its replica nodes. In order
to evaluate the expected size of the mutually connected component, we average the messages over this ensemble of the
network of networks. We indicate with σα the average message within a layer ⟨σiα→jα ⟩ = σα .
The equations for the average messages σα are given in terms of the parameter p and the generating functions Gα0 (z ) and
α
G1 (z ) defined as
Gα0 =

Pα (k)z k
k
 kPα (k)
Gα1 = z k−1 , (129)
k
⟨k⟩α
where ⟨k⟩α indicates the average degree in layer α , i.e. ⟨k⟩α = k kPα (k). We assume here that the damage of the nodes in

the different layers is correlated, and that therefore a node (i, α) is damaged with probability p together with all its replica
nodes. We have
 
M
 M
 M

P ({si,a }) = p δ(1, si,α ) + (1 − p) δ(0, si,α ) , (130)
i =1 α=1 α=1
where δ(x, y) indicates the Kronecker delta. This assumption is needed if we interpret the set of different replica nodes (i, α)
in a multiplex like a single node having different types of interactions. Therefore, using Eq. (127), in Ref. [171] it was found
that
 
β
  
σα = p 1 − G0 (1 − σβ ) 1 − Gα1 (1 − σα ) .

(131)
β̸=α
Moreover, using Eq. (128), the probability for Sα = ⟨Siα ⟩ = S that a randomly chosen node belongs to the mutually
connected component is
M 
β
 
S=p 1 − G0 (1 − σβ ) . (132)
β=1
Equivalent equations have been previously found in Refs. [25,153] using an alternative derivation. In particular, if all layers
β have the same topology, then the average messages within different layers β are all equal σβ = σ , and are given by
σ = p [1 − G0 (1 − σ )]M −1 [1 − G1 (1 − σ )]. (133)
Furthermore, the probability S = ⟨Siα ⟩ that a node in a given layer is in the mutually connected component is given by
S = p [1 − G0 (1 − σ )]M . (134)
In Ref. [170] the case of targeted attack toward high or low degree nodes have been considered. It has been found in this
context that in contrast with the single-layer network scenario, SF interdependent networks are difficult to defend using
the strategy of protecting high degree nodes.
4.3.1. Case of a multiplex formed by M Poisson networks with the same average degree
In the case of a multiplex formed by M Poisson networks with the same average degree ⟨k⟩α = c, we have that the
equation for the order parameter becomes
M
σ = S = p 1 − e−cS .

(135)
For every finite M ≥ 2 it is possible to show that the MCGC emerges discontinuously at p = pc (M ). The equation for the
order parameter σ can be written as
M
gy=cp (Sc = x) = x − y 1 − e−x = 0.

(136)
0.5
gy x
0
y=2.2
y=2.455
y=2.6
0.5
0 1. 2. 3.
x
Fig. 20. (Color online) The function gcp=y (x = Sc ) = x − y(1 − e−x )2 plotted for different values of y = cp. The solution of the equation gcp (Sc ) = 0
determines the fraction S of nodes in the mutually connected component of a duplex formed by two Poisson networks with average degree c, when a
fraction 1 − p of nodes has been damaged. As cp < 2.45541 . . ., only the trivial solution S = 0 exists, while at cp = 2.45541 . . . a non trivial solution S > 0
emerges discontinuously.
3
Sc
0
0 1 2 3 4 5
cp
Fig. 21. The quantity Sc as a function of the product cp, where S is the fraction of nodes in the mutually connected component of a duplex network formed
by two Poisson networks with average degree c, and 1 − p is the fraction of nodes initially damaged in the multiplex network.
The equation gy (x) = 0 has always a solution σ = 0, but at y = cp = yc another solution appears discontinuously when
the function gy (x) is tangential to the x axis at the point x = xc , y = yc . Therefore, in order to find the critical values x = xc
and y = yc , we can impose the set of equations
gy (x) = 0
dgy (x)
= 0. (137)
dx
In Fig. 20 we plot the function gy (x) for M = 2, showing that at y = yc = 2.45541 . . . the equation gy (x) develops a non
trivial solution at Sc = xc = 1.25643 . . ..
In Fig. 21 we plot the order parameter S as a function of y = cp, for the case of a duplex, i.e. M = 2. It is clear that the order
parameter has a discontinuity at y = yc = 2.45541 . . .. Moreover, it can be shown that the order parameter has a singularity
for y → y+ c . In fact, it has been shown [35,168] that the transition is a discontinuous hybrid transition characterized by a
square-root singularity at the phase transition, in other words the order parameter, close to the transition, scales as
S − Sc ∝ (cp − yc )1/2 , (138)

for cp − yc = ϵ and 0 < ϵ ≪ 1.
4.3.2. Case of a duplex network formed by two Poisson networks with different average degree
In Ref. [171], the case of a duplex network formed by two Poisson networks (network A and network B) with different
degree distribution (with average degree respectively ⟨k⟩A = zA and ⟨k⟩B = zB ) has been studied. The equations determining
the emergence of the MCGC reduce to a single equation for S = σA = σB given by
S = p(1 − e−zA S )(1 − e−zB S ). (139)

a b 10
0.5
8 0.4
0.3
Sc
0.2
0.8 6
0.1
ZB
5 0.6 0.0
0.0 0.2 0.4 0.6 0.8 1.0
4
4 ZA/ZB
0.4 S percolating
ZB 3 0.2 2
2 0.0
5
4 0
3 0 2 4 6 8 10
2 ZA
ZA
Fig. 22. Phase diagram (for p = 1) of a multiplex formed by two Poisson networks (network A and network B) with average degrees ⟨k⟩A = zA and
⟨k⟩B = zB respectively.
Source: Reprinted figure from Ref. [171]. © IOP Publishing Ltd and Sissa Medialab srl. Reproduced by permission of IOP Publishing. All rights reserved.
Therefore, the phase diagram only depends on the parameters (pzA , pzB ). Also in this case, the emergence of the mutually
connected component is described by a discontinuous hybrid transition, at the points of the phase diagram satisfying
simultaneously the following set of equations
h(S ) = S − p(1 − e−zA S )(1 − e−zB S ) = 0
h′ (S ) = 0. (140)
The phase diagram of the model is shown in Fig. 22 for p = 1.
4.3.3. Multiplex networks formed by layers of scale-free networks

In the case of single-layer networks, SF networks (with power-law degree distribution P (k) = Ck−γ and γ ∈ (2, 3])
are known to be much more robust to random failures than networks with a finite second moment ⟨k2 ⟩, e.g. of Poisson
networks [157–159]. Indeed, for SF networks of size N, the giant component emerges if the fraction of nodes that are not
initially damaged p > pc with pc → 0 as N → ∞. Therefore, these networks continue to have the giant component also
if most of the nodes are initially damaged. A finite percolation threshold in these networks is a finite size effect, i.e. the
percolation threshold disappears in the limit N → ∞. The situation is completely different for multiplex networks formed
by interdependent SF networks [25]. The percolation threshold p = pc for the mutually connected component is finite also
for multiplex networks formed by layers of SF networks [25]. In the case in which the layers of the multiplex have the same
power-law degree distribution p(k) = Ck−γ , the mutually connected component emerges discontinuously for p = pc as
a hybrid phase transition with a square root singularity [25,168]. In these networks, pc is not vanishing but remains finite
as in the case of a multiplex in which the layers are formed by Poisson networks [25,168]. In Fig. 23 we show the size of
the mutually connected component for two SF networks with the same power-law exponent γ . As γ → 2, and keeping
the minimal degree constant, the discontinuity of S becomes smaller and smaller but is not vanishing as long as γ > 2. In
Ref. [25] it was shown that actually another major difference exists between SF single-layer and multiplex networks. In fact,
if we consider a multiplex formed by two SF networks with the same degree distribution, and compare their percolation
threshold pc keeping the average degree ⟨k⟩ constant, changing only the power-law exponent we obtain that the percolation
threshold p = pc (γ ) increases as γ is decreased. This implies that multiplex formed by broader SF are more fragile than
multiplex with a steeper degree distribution (see Fig. 24). This surprising result shows that the robustness of multiplex
networks has very new and unexpected features.
4.3.4. Cascading failures

In the case of a multiplex network, one can follow the propagation of the failures back and forth from one network to the
others [25,155,176]. In the specific case of a duplex, it is possible to characterize how the failure in one network propagates
(0)
in the other layer in the following way. We start form the layer A and we consider the equations determining the size SA
of the giant component when the random damage to a fraction 1 − p of the nodes has been inflicted to the network. In
(0) (0)
particular, SA will depend on the average message σA indicating what is the probability that a node at an end of a link in
layer A is in the giant component of the same layer. Then, we consider layer B, where we assume that all the nodes whose
replica nodes in layer A are not in the giant component of layer A are damaged. Therefore, we consider all the nodes in layer B
that remain in the giant component of layer B after this damage propagating from layer A has been inflicted. The probability
1.0
100
0.8 10-4
0.6 10-8
0 p 1
0.4
0.2
0.0
0.0 0.2 0.4 0.6 0.8 1.0
p
Fig. 23. (Color online) Fraction of nodes S in the MCGC as a function of p for two symmetric power-law distributed networks with, from right to left,
power-law exponent γ = 2.8, 2.5, and γ = 2.1 and same minimal degree. The height of the jump becomes very small as the power-law exponent γ
approaches 2, but is not zero, as seen in the inset, which show S vs. p on a logarithmic vertical scale for γ = 2.1.
(0) (1)
that the nodes are not damaged is p[1 − G0 (1 − σA )] and, using this probability one can calculate the probability σB that
(1)
a random link in layer B reaches a node in the giant component of layer B and the probability SB that a random node in
layer B is in the remaining giant component of layer B. Iterating this process back and forth from one layer to the other, it
is possible to describe the propagation of cascading events in the interconnected network. At the end of this iterative pro-
(n) (n)
cedure, the remaining nodes are the nodes in the mutually connected component. We will call σA and σB the probability
that, in following a link either in layer A or in layer B, one reaches a node in the giant component of layer A or in the giant
(n) (n)
component of layer B at the nth step of this cascading process. Similarly, we will call by SA and SB the probability that a
node in layer A or in layer B is in the giant component of layer A or of layer B. The propagation of failures from one network
(n) (n)
to the other can be described by the following equations for the probabilities σA , σB
σA(0) = p[1 − GA1 (1 − σA(0) )]

σB(1) = p[1 − GA0 (1 − σA(0) )][1 − GB1 (1 − σB(1) )]
...
σA(2m+2) = p[1 − GB0 (1 − σB(2m+1) )][1 − GA1 (1 − σA(2m+2) )]
σB(2m+3) = p[1 − GA0 (1 − σA(2m+2) )][1 − GB1 (1 − σB(2m+3) )], (141)
(n) (n)
and by the following equations for the probabilities SA , SB
(0)
SA = p[1 − GA0 (1 − σA(0) )]
(1)
SB = p[1 − GA0 (1 − σA(0) )][1 − GB0 (1 − σB(1) )]
...
(2m+2)
SA = p[1 − GB0 (1 − σB(2m+1) )][1 − GA0 (1 − σA(2m+2) )]
(2m+3)
SB = p[1 − GA0 (1 − σA(2m+2) )][1 − GB0 (1 − σB(2m+3) )]. (142)
See Refs. [155,176] for the general equations describing the cascades of failures in multiplex networks with a generic value
of M. Eventually, these cascading events stop when the damage does not propagate further in the duplex network. Close
to the phase transition p ≃ pc , the cascades become systemic events and propagate until the mutually connected compo-
nent completely disappears. This phenomenon reveals that the interdependent networks can be much more fragile than
single-layer networks and are prone to abrupt cascading failures.
In Ref. [25] the size of the network after n steps of propagation of the failures from one network to the other was first
studied and characterized. In particular, pn indicates the fraction of non-damaged nodes in the network after n steps. In
other words
(n)
for n = 2m, m = 1, 2, . . .

SA
pn = (n) (143)
SB for n = 2m + 1, m = 1, 2, . . . .
In Fig. 25 the subsequent values of pn are plotted for different realizations of a duplex Poisson network with p = 2.45/⟨k⟩ <
pc . The solid red line indicates the theoretical predictions valid in the limit N → ∞. These predictions are very good for
1.0
0.8
0.6
0.4
0.2
0.0
0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90
p
Fig. 24. (Color online) Simulation results for the fraction of nodes p∞ = S in the mutually connected component as a function of p for SF networks with
power-law exponent λ = 3, 2.7, 2.3, an ER network and a Random Regular (RR) network, all with an average degree ⟨k⟩ = 4. The simulation results are
performed at finite number of nodes N, and therefore are not clearly discontinuous due to finite size effects. Nevertheless, it can be seen already that pc is
higher for broader distributions.
Fig. 25. (Color online) The fraction of nodes pn in the mutually connected component remaining after n steps of the propagation of the damage back
and forth between the two layers of a duplex network. In this case, the duplex has two layers with Poisson degree distribution and the same average ⟨k⟩.
Moreover, p is set at the value p = 2.455/⟨k⟩ < pc , and the total number of nodes N is given by N = 12800. The red circles indicate the theoretical
predictions valid in the limit N → ∞.
small n, but for large values of n the simulations start to show finite size fluctuations. In particular, the simulations result
either in a multiplex network with a non zero mutually connected giant component or in a multiplex network without it,
while the deterministic theory predicts that in the N → ∞ the mutually giant component disappears in the network. In
Ref. [25] an analysis of the finite size effects was first performed, yielding the following scenario for the average number of
steps ⟨n⟩ of a cascade of failures as a function of p and N:
• Case p < pc . The average number of steps ⟨n⟩ of the cascade of failures can be approximated by
1
⟨n⟩ ∼ √ . (144)
pc − p
• Case p = pc . The probability P (n) that a critical avalanche of failures has n steps scales like
n
−α
e N 1/4
P (n) = C , (145)
N 1/4
where C > 0 and α > 0 are constants, independently on the number of nodes N in the multiplex. Therefore, the average
number of steps of a critical avalanche of failures ⟨n⟩ is diverging with the multiplex size N as
⟨n⟩ ∼ N 1/4 . (146)

• Case p > pc . The average number of steps ⟨n⟩ of the cascade of failures can be approximated by
ln N
⟨n⟩ ∼ √ . (147)
p − pc
4.3.5. Partial interdependence

If the networks are only partially interdependent, the transition becomes continuous [35,171]. Different versions of
this model have been proposed modulating the degree distribution, and the correlations between the layers [177–180].
Here we show how a passage between a discontinuous and a continuous transition can occur in the simplest setting of a
multiplex, and in particular in a duplex, formed by ER networks. Let us consider a multiplex of M layers where each node
(i, α) is interdependent on each of its replica nodes (i, β) with probability r. In this case, the message passing Eqs. (127)
still remain valid, where Ci (α) defines all the set of layers β for which the node (i, α) is interdependent on its replica node
(i, β). We assume that every layer α has degree distribution Pα (k), and that every node (i, α) is interdependent on every
one of its replica nodes. We will describe the emergence of the mutually connected component by averaging the messages
over links in each layer. We indicate with σα the average message within a layer ⟨σiα→jα ⟩ = σα . Similarly to the case of a
fully interdependent multiplex, we will assume that the damage of the nodes in the different layers is correlated and that,
therefore, we have P ({siα }) given by Eq. (130) that we rewrite here for convenience, i.e.
 
M
 M
 M

P ({si,a }) = p δ(1, si,α ) + (1 − p) δ(0, si,α ) . (148)
i =1 α=1 α=1
The equations for the average messages within a layer are given in terms of the parameter p and the generating functions
Gα0 (z ) and Gα1 (z ) defined in Eq. (129). Using Eq. (127) we find
 
β
  
σα = p 1 − rG0 (1 − σβ ) 1 − Gα1 (1 − σα ) .

(149)
β̸=α
Moreover, using Eq. (128) we can derive the probability for Sα = ⟨Siα ⟩ = S that a randomly chosen node belongs to the
mutually connected component
 
β
  
1 − rG0 (1 − σβ ) 1 − Gα0 (1 − σα ) .

S=p (150)
β̸=α
In particular, if all layers β have the same topology, then the average messages within different layers are all equal σβ = σ ,
and are given by
σ = p [1 − rG0 (1 − σ )]M −1 [1 − G1 (1 − σ )]. (151)

S = p[1 − rG0 (1 − σ )]M −1 [1 − G0 (1 − σ )] . (152)

For r = 0 the nodes of each layer have no interdependencies with the nodes of the other layers, therefore each layer will
percolate independently, and the percolation transition will be a second order continuous phase transition. On the contrary,
when r = 1 we have complete interdependency of each node of a given layer on all its replica nodes in the other M − 1
layers, and the transition will be a discontinuous hybrid transition. Fig. 26 reports simulations results of these two limits,
in duplex networks of different degree distributions. In particular, there is a tricritical point (r = rc , p = pc ) for which the
transition changes its nature, from discontinuous (for r > rc ) to continuous (for r < rc ).
Let consider a duplex network formed by two Poisson networks with the same average degree ⟨k⟩ = c. The order
parameter σ = S is the solution of the equation
S = p 1 − re−cS 1 − e−cS .
  
(153)
The points of the hybrid phase transition can be found by imposing

h(S ) = S − p 1 − re−cS 1 − e−cS = 0,
  
dh(S )
= 0, (154)
dS
Fig. 26. (Color online) Simulation results for the fraction of nodes β∞ = S in the MCGC of a duplex formed by N = 50 000 nodes, as a function of p.
For strong coupling (high value of r) between the networks, there is a discontinuity in β∞ (Poisson networks—circle, and SF networks—up rectangle).
For weak coupling (low value of r) between the networks, the change in β∞ is described by a continuous transition (Poisson networks—square, and SF
networks—right rectangle).
and by looking for a non trivial solution S > 0. This solution can be found for any r > rc characterizing the level of partial
interdependence of the tricritical point of the model. For r < rc , the emergence of the mutually connected component
is dictated by a second order phase transition at a critical value of p that can be obtained by imposing the conditions
h(0) = 0, h′ (0) = 0. The tricritical point separates these two regimes, and results [171] from imposing h(0) = 0, h′ (0) = 0,
and h′′ (0) = 0. This system of equations yields the parameters of the tricritical point:
1
rc = ,
3
3
cpc = . (155)
2
The avalanches of failures in the partially interdependent networks have been characterized in Refs. [35,177–179], where
the Authors observe a different scenario depending on the nature of the phase transition in the network. In particular, it has
been observed, that if the system is close to the hybrid discontinuous phase transition, the avalanches of failures are very
long. In fact, the order parameter pn decays very slowly with the number of steps n of the propagation of the damage back
and forth in the two layers, and finally decays to a zero value. On the contrary, the critical avalanches of failures close to the
second order phase transition are characterized by an order parameter pn that decays very rapidly with the number of steps
n of the propagation of the damage.
4.3.6. Percolation in multiplex networks with multiple support-dependence relations

In Ref. [169], the case of a duplex is considered in which each node (i, α) can be at the same time connected to several
nodes (j, β) in layer β ̸= α and can have multiple connections within the same layer α . It is assumed that a node (i, α) is in
the mutually connected component if the following two conditions are met:
(a) there is at least one neighbor (j, α) of node (i, α) in layer α that is in the mutually connected component;
(b) for every layer β ̸= α , there is at least one node (j, β) connected to node (i, α) which is in the mutually connected
component.
αβ
Assuming that the layers are two (layer A and layer B) and that the degrees ki are drawn from a Poisson degree distribution
αβ A ,A
Pαβ (k ) with ⟨k ⟩ = a, ⟨k ⟩ = b, ⟨k ⟩ = ã and ⟨k ⟩ = b̃, it has been shown in Ref. [169] that the equations determining
BB AB BA
the probability σA (σB ) that a node in layer A (B) is in the mutually connected component are given by
 
σA = pA 1 − e−ãσB 1 − e−aσA ,

 
σB = pB 1 − e−b̃σA 1 − e−bσB ,

(156)
where 1 − pA (1 − pB ) is the initial fraction of damaged node in layer A (B). For these networks, the percolation is hybrid and
discontinuous, but in the limit ã, b̃ → ∞ it becomes continuous.
4.4. Interdependent networks of networks
Following Refs. [26,152–156], we will consider three major types of networks of networks.
• Case I: Networks of networks with fixed supernetwork and interdependent links allowed only between replica nodes, where
every node (i, α) of layer α is interdependent on its replica node (i, β) in the same set of layers β [152]. This ensemble
of network of networks has been described in Section 3.2.1.
• Case II: Networks of networks with fixed supradegree distribution, where every node (i, α) is interdependent on q = qα
replica nodes (i, β) of randomly chosen layers β [154]. This ensemble has been described in Section 3.2.2.
• Case III: networks of networks with fixed supernetwork and random permutation of the labels of the nodes, where every node
(i, α) is interdependent on a node (πα,β (i), β) if layer α is interdependent on layer β , where πα,β is a random permutation
of the labels {i} of the nodes [26,153,155,156].
4.4.1. Case I: networks of networks with fixed supernetwork and interdependent links allowed only between replica nodes
In this subsection, we will consider networks of networks in which, if network α is interdependent with network β , each
node (i, α) of network α is interdependent on node (i, β) of network β , and vice versa. This ensemble has been introduced
in Section 3.2.1, and it has been graphically represented in Fig. 15. The corresponding supra-adjacency matrix has elements
aiα,jβ = 1 if there is a link between node (i, α) and node (j, β), and zero otherwise. In these specific networks, aiα,jβ = 0
if both i ̸= j and α ̸= β . The graph of interdependencies is characterized by an adjacency matrix Aαβ such that, for every i,
Aαβ = aiα,iβ . The matrix Aαβ is the adjacency matrix of the supernetwork G.
In Ref. [152] it has been shown that for this type of networks of networks, where each node can be interdependent only
on its replica nodes in the other layers, the supernetwork can be a tree or can contain loops, and in both cases, as long as the
supernetwork is connected, the size of the mutually connected component is determined by the same equations determining
the MCGC in a multiplex network formed by the same layers. This result is strongly dependent on the type of considered
networks of networks. For example, for networks of networks of case III the presence of loops in the supernetwork change
significantly the mutually connected component [26,153,155,156].
In Ref. [152] it has been shown that the problem of a wide range of networks of networks can be actually reduced to the
well studied problem of multiplex networks. The main technical results of Ref. [152] has been to show that the messages
σiα→jα within each layer are given by Eq. (127), that we rewrite here for convenience
    
    
σiα→jα = siα 1 − (1 − σℓα→iα ) siβ 1 − (1 − σℓβ→iβ ) , (157)
ℓ∈Nα (i)\j β∈Ci (α)\α ℓ∈Nβ (i)
 
where Ci (α) is the set of layers β such that the ‘‘replica nodes’’ (i, β) are connected to the node (i, α) by paths going through
the interdependency links, and where Nα (i) = ℓ indicates that set of label ℓ of nodes (ℓ, α) that are neighbors of node (i, α)
in layer α . Ref. [152] considered the case in which the supernetwork is given (fixed by its adjacency matrix Aαβ ), and where
each layer α is generated independently from a configuration model with a degree distribution Pα (k). The supernetwork is,
therefore, not random, while the layers are infinite uncorrelated random networks. It is assumed that each layer α has a
degree sequence {kαi }, and that the degrees of the replica nodes in the different layers are uncorrelated. Furthermore, it is
assumed that the nodes (i, α) are removed from layers with probability 1 − pα (i.e. they are instigators of cascading failures).
Consequently, the probability P ({siα }) is given by
M 
 N
P ({siα }) = psαiα (1 − pα )1−siα . (158)
α=1 i=1
In order to evaluate the expected size of the mutually connected component, the messages are averaged over this ensemble.
If σα indicates the average message within a layer ⟨σiα→jα ⟩ = σα , the equations for the average messages within a layer are
β β
given in terms of the parameters pβ = ⟨siβ ⟩ and the generating functions G0 (z ) and G1 (z ) are defined as
β

G0 = Pβ (k)z k ,
k
β
 kPβ (k)
G1 = z k−1 . (159)
k
⟨k⟩β
Without loss of generality, we can consider a connected supernetwork G formed by M layers. In fact, if the supernetwork is
not connected, we can consider each connected component of the supernetwork separately. Using Eq. (157), the expression
for σα is given by
β
 
σα = pα pβ [1 − G0 (1 − σβ )] [1 − Gα1 (1 − σα )]. (160)
β∈G\α
Moreover, using Eq. (128), the probability Sα = ⟨Siα ⟩ that a randomly chosen node in layer α belongs to the mutually
connected component is given by
β
 
Sα = pβ [1 − G0 (1 − σβ )] . (161)
β∈G
Thus, as long as the supernetwork is connected, the resulting mutual component is determined by the same equations
governing the case of a mutually connected component between the same layers, where the only difference between Eqs.
(160) and (131) and between Eqs. (161) and (132) depends on the different choice of the probability P ({siα }) for the initial
damage inflicted in the network. In the case in which all layers β have the same topology and equal probabilities pβ = p,
the average messages within different layers are all equal σβ = σ , and are given by
σ = pM [1 − G0 (1 − σ )]M −1 [1 − G1 (1 − σ )]. (162)
S = {p[1 − G0 (1 − σ )]}M . (163)
Therefore, the size of the mutually connected component always depends on the number of layers M in the connected
supernetwork.
When we include partial interdependence between the nodes, by introducing the probability r that a node is
interdependent on a linked node in another layer, the Eqs. (160)–(161) become respectively
β
 
σα = pα pβ [1 − rG0 (1 − σβ )] [1 − Gα1 (1 − σα )], (164)
β∈G\α
and
β
  
Sα = pβ [1 − rG0 (1 − σβ )] [1 − Gα0 (1 − σα )]. (165)
β∈G\α
From these results, four main conclusions can be drawn:

(a) the transition is hybrid and discontinuous as soon as M ≥ 2,
(b) the critical value pc always depends on the number of layers M,
(c) when the mutually connected component emerges at p = pc , every layer of the network of networks contains a finite
fraction of nodes that is in the mutually connected component, i.e. in each layer the percolation cluster emerges at the
same time at p = pc ,
(d) if we include a partial interdependence of the nodes, the transition changes from discontinuous to continuous at a
tricritical point, but also in this case either all layers percolate, or any of them does.
4.4.2. Case II: networks of networks with given superdegree of the layers
Networks of networks can have a structure more random than the one considered in the above subsection. In Ref. [154],
the attention has been addressed to the case in which every node of a network (layer) α is connected with qα > 0 randomly
chosen replicas in some other networks, and it is interdependent of these nodes with probability r. Each network (layer) α is
generated from a configuration model with the same degree distribution Pα (k) = P (k), and each node (i, α) is connected to
qα other ‘‘replica’’ nodes (i, β) chosen uniformly at random. We further assume that the degree sequence in each layer is {kαi }
and that the degrees of the replicas of node i are uncorrelated. This is the ensemble described in Section 3.2.2, and graphically
represented in Fig. 16. Every network of networks in this ensemble has a supra-adjacency matrix a with probability P (a)
given by
    
M 
N M 
N 
α
   
P (a) = δ ki , δ qα , δ aiα,jβ , 0 .
 
aiα,jα aiα,iβ (166)
α=1 i=1 j̸=i β̸=α α=1 i=1 β̸=α j̸=i
The initially inflicted random damage is defined by the variables siα . In particular, if siα = 0, the node (i, α) is initially
damaged, otherwise siα = 1. The following probability P ({siα }) is considered:
M 
 N
P ({siα }) = psiα (1 − p)1−siα . (167)
α=1 i=1
In order to quantify the expected size of the mutually connected component in this ensemble, Ref. [154] considers the
average of the messages over this ensemble. The message passage equations for this problem are given by Eqs. (127). Let us
first consider the case r = 1, where all the links between the different layers indicate an interdependency between the two
0.6 3
0.4 2
qmax
0.2 1
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
p p
Fig. 27. Plots of the order parameter Σ (left panel) and of the maximal superdegree of percolating layers qmax (right panel) vs. p for a configuration model
with a Poisson P (k) distribution with average ⟨k⟩ = c = 20 and a SF P (q) distribution with γ = 2.8, and with minimal degree m = 1 and maximal
superdegree Q = 103 .
linked replica nodes. The equations for the average message within a layer are given in terms of the parameter p = ⟨siα ⟩,
q q
and the generating functions Gk0 (z ), Gk1 (z ), G0 (z ), G1 (z ) are given by
  kP (k)
Gk0 (z ) = P (k)z k , Gk1 (z ) = z k −1 ,
k k
⟨k⟩
q
 q
 qP (q)
G0 (z ) = P (q)z q , G1 (z ) = z q −1 . (168)
q q
⟨q⟩
In particular, if σq indicates the average messages within a layer α of degree qα = q, we obtain
 s−1
  q′ P (q′ )
σq = p P (s|q) p[1 − G0 (1 − σq′ )] [1 − G1 (1 − σq )], (169)
s q′
⟨q⟩
where P (s|q) indicates the probability that a node i in layer α with qα = q is in a connected component C (i, α) of the local
supernetwork of cardinality (number of nodes) |C (i, α)| = s.
Similarly, the probability that a node i in a layer α with superdegree qα = q is in the mutually connected component
Sq = ⟨Siα ⟩ is given by
 s−1
  q′ P (q′ )
Sq = p P (s|q) p[1 − G0 (1 − σq′ )] [1 − G0 (1 − σq )]. (170)
s q′
⟨q⟩
In Ref. [154] it has been shown that Eqs. (169) and (170) can be expressed in the following simplified way:
σq = p(Σ )q [1 − Gk1 (1 − σq )],
Sq = p(Σ )q [1 − Gk0 (1 − σq )],
 
 q′ P (q′ )  q′ P (q′ ) ′
Σ= p [1 − Gk0 (1 − σq′ )] (Σ )q −1 . (171)
q′
⟨q⟩ q′
⟨q⟩
Therefore, in this ensemble the parameter Σ determines both σq and Sq for any value of the superdegree q. For this reason,
Σ can be considered the order parameter.
From the study of these equations, the following scenario can be drawn in the limit M → ∞:
(a) if the degrees qα are heterogeneous, the percolation transitions are multiple, each one corresponding to the emergence
of a percolation cluster in each layer with a different value of q;
(b) each of these transitions is hybrid and discontinuous, and
(c) the layers with higher number of interdependencies are more fragile than those with a smaller number of interdepen-
dencies.
The increased fragility of those layers with higher superdegree is in sharp contrast with what happens in the percolation of
single layers, in which nodes of high degree are the more robust. In Fig. 27, the order parameter Σ is shown for a power-
law distribution P (q) ∼ q−γ , with γ = 2.8 and a Poisson distribution P (k) with average ⟨k⟩ = c = 20. The multiple
0.5
0.4 q=1
q=2
0.3 q=3
σq
0.2
0.1
0
0 0.2 0.4 0.6 0.8 1
p
Fig. 28. (Color online) Plot of σq = Sq (fraction of nodes in a layer of superdegree q, belonging to the mutual component) vs. p for different values
q = 1, 2, 3 in the configuration model of the network of networks having a Poisson P (k) distribution with average ⟨k⟩ = c = 20 and a SF P (q) distribution
with γ = 2.8, minimal degree m = 1 and maximal superdegree Q = 103 . For each value of q = 1, 2, 3, σq emerges discontinuously, with a jump, which
becomes smaller and smaller with increasing q. The emergence of σ2 is accompanied by a discontinuity of σ1 (p). The emergence of σ3 is accompanied by
discontinuities of σ1 (p) and σ2 (p).
discontinuous phase transitions are apparent. Moreover, the degree qmax corresponding to the degree of interdependencies
of the percolating layers with maximal value of qα is shown. Each transition, characterized by a discontinuity of the order
parameter Σ , corresponds to the activation of the layers with increasing superdegree qα . Fig. 28 shows that the size of the
percolation cluster σq in the layers with q number of interdependencies have multiple discontinuities in correspondence
with the discontinuities of the order parameter Σ . Moreover, it is shown that the percolating cluster of layers with higher
superdegree q is disrupted before the percolating cluster of layers with smaller superdegree.
For r < 1, Eqs. (171) are modified as follows:
σq = p(r Σ + 1 − r )q [1 − Gk1 (1 − σq )],
Sq = p(r Σ + 1 − r )q [1 − Gk0 (1 − σq )],
 
 q′ P (q′ )  q′ P (q′ ) ′
Σ= p[1 − Gk0 (1 − σq′ )] (r Σ + 1 − r )q −1 . (172)
q′
⟨q⟩ q′
⟨q⟩
We note here that, in the case in which the local supernetwork is regular, i.e. P (q) = δ(q, m), and each layer is formed by a
Poisson network with ⟨k⟩ = c, one finds Σ [r Σ + 1 − r ] = σm = σ satisfying the following equation
 
1 m
(1 − e−c σ ).

σ =p 1−r + (1 − r ) + 4r σ
2 (173)
2
The emergence of the mutually connected component in the configuration model of a network of networks with r < 1
can always display multiple percolation transitions corresponding to the activation of layers in increasing value of qα .
Nevertheless, these transitions can be either continuous or discontinuous, depending on the value of r.
4.4.3. Case III: networks of networks with fixed supernetwork and random permutation of the labels of the nodes
Here we consider the robustness of networks of networks described in Section 3.2.3 where every node (i, α) is
interdependent on a node (πα,β (i), β) if layer α is interdependent on layer β , and πα,β is a random permutation of the
labels {i} of the nodes.
In Refs. [26,153,155,156] this structure for the network of networks is implicitly assumed and they demonstrate that
the structure of the supernetwork matters. In fact, if the supernetwork G is a tree, then the size of its mutual component is
determined by the same equations that determine the size of the MCGC in a network of networks of case I. On the other hand,
it was found that, for a supernetwork with loops, this is not any more true [26,155,156]. In particular, in Refs. [26,155,156]
the size of the MCGC for a network of networks was explicitly derived for the case of a regular random supernetwork. The
size of the mutually connected component when each layer is a Poisson network of average degree c and the supernetwork
is regular (i.e. the superdegree q of each layer has distribution P (q) = δ(q, m)) is given by
 
1 m
(1 − e−c σ ),

σ =p 1−r + (1 − r ) + 4r σ
2 (174)
2
where here partial interdependency of the links is taken into account and r indicates the probability that a node is
interdependent on a linked node in another layer [26,156]. We observe here that the Eq. (174) is equivalent to Eq. (173)
1
MP
0.8 UC
MN
0.6
M
0.4
0.2
0
0 0.5 1 1.5 2 2.5 3 3.5 4
mean degree, z
Fig. 29. (Color online) The mutually connected component for a duplex of Poisson networks with different level of degree correlations. The symbol MP
indicates maximally-positive degree correlations, the symbol MN indicates maximally-negative degree correlations, the symbol UC indicates no degree
correlations.
derived for networks of networks belonging to case II. This result can be generalized. In fact if the supernetwork is a random
network with given superdegree distribution the mutually connected component of a network of network of case III follows
the same equations determining the MCGC in the network of network of case II. Therefore the following general scenario
holds for networks of networks in case III:
• If the supernetwork is a tree, the MCGC is determined by the same equations determining the MCGC in a network of
networks of case I. Therefore, in this case either all the layers are percolating or none does.
• If the supernetwork is a random network with given superdegree distribution, the MCGC is determined by the same
equations determining the MCGC in a network of networks of case II. Therefore, in this case, there are multiple phase
transitions and layers with higher superdegree are more fragile than layers with smaller superdegree.
4.5. Effects of correlations and embedding space on percolation properties of multilayer networks
The correlations on multiplex networks have important consequences on their percolation properties. Here we will
consider the effects of overlap of the links, degree correlations, and those of the embedding space.
4.5.1. Percolation in networks with overlap of the links

The overlap of the links can change significantly the properties of the percolation transition. In fact, in the limit in which
we have a multiplex of M totally overlapping layers, the mutually connected giant component of the multiplex is identical
to the giant component of a single layer, and the percolation transition is of the second order.
For these reasons, it would be natural to expect that, when the overlap between the layers is over a threshold value, the
emergence of the mutually connected component is continuous, while when the overlap is below such a critical value, the
emergence of the mutually connected component is described by a hybrid transition.
This problem was tackled by several groups [48,51,181,182], finding different results. In Ref. [51], the Authors have
studied the phase diagram corresponding to the message passing algorithm finding a tricritical point for a non vanishing
value of the overlap of the links. In [181] the mutually connected component has been studied with different analytical
techniques. In this latter case, the Authors do not find any tricritical point, and the percolation transition remains
discontinuous for any finite fraction of links without overlap.
4.5.2. Percolation in networks with degree correlations

The presence of degree correlations in multiplex networks can modify drastically the percolation threshold pc
[183,184]. The general equations for the percolation transition in correlated multiplex have been first derived in Ref. [168],
while Refs. [48,183,184] considered the case of single-layer networks with the help of the message passing algorithm of
Ref. [175]. Ref. [183] considered multiplex networks in which the replica nodes have the same degree. Refs. [184,185] studied
the robustness of the interdependent world-wide airport network and the world-wide port network with positive degree
correlations and clustering. Finally, in Ref. [48] different types of correlation between the degree of the nodes in the different
layers have been considered. The Authors investigated the maximal changes induced by the degree correlations for duplex
with the same degree sequences, but either uncorrelated by the degree or maximally-positive correlated or maximally-
negative correlated. For maximally-positive correlated multiplex networks, the percolation threshold pc is smaller than
that for the maximally-negative correlated case. The networks with uncorrelated degrees in the different layers have a
percolation threshold pc that is in between the previous two. In Fig. 29 the dependence of the fraction of nodes in the MCGC
of correlated Poisson networks is shown as a function of the average degree of the nodes. In Ref. [48], the role of targeted
attack of multiplex networks is also investigated and discussed.
Other papers have investigated the effect of the degree correlations within a layer on the percolation properties of the
network, and in particular assortativity [185], or the effect of clustering [186]. Finally, in Ref. [187], the effect of degree
correlations has been studied on a duplex network with partial interdependence.
4.5.3. Percolation in spatial multiplex networks

Many networks are not only interdependent, but also embedded in a physical space. For instance, infrastructures are
mostly spatial and therefore it is essential to investigate the role of space in determining their robustness. Ref. [165] has
shown that when the multiplex networks are embedded in real space, the interdependencies do not always sharpen the
percolation transition. The topic is the subject of debate in the literature [166,167]. Moreover, the nature of the percolation
transition can depend on the typical distance in space of the interdependent nodes. In fact, by modulating the typical distance
from two interdependent nodes, the transition can change from discontinuous (for large typical distances), to continuous
(for typical distances lower than a threshold value) [188]. This model has been extended in several directions [189–191].
For example, in Ref. [189] partial interdependence was considered, and the Authors found that, in contrast to unembedded
networks, for any fraction of dependency links the system collapses in an abrupt transition. Reference [190] tackled
the problem of a network that recovers after each cascade. Furthermore, Ref. [191] studied the case of localized attack
percolation in spatial networks.
4.6. Other percolation problems
4.6.1. Classical percolation

The percolation of multilayer networks can also be studied in the absence of interdependencies. This is the case in which
all the connections, either those between nodes in the same layer and those between nodes in different layers, have the same
role and all the links are ‘‘dependency links’’. In particular, one can consider the giant component as a function of the fraction
of damaged nodes in the multilayer structure. In the context of multilayer transportation systems, involving different means
of communication such as buses, undergrounds, trains, etc., one station will be in the giant component if by following at
least a path using any number of transportation systems, it is possible to reach a finite fraction of other stations. In this case,
a node of the structure belongs to the giant component of the multiplex if at least one of its links connects the node to the
giant component. This problem has been studied in Ref. [80], considering the model of networks of networks described in
Section 3.2.4, and graphically shown in Fig. 18. Previously similar problems have been already addressed in the context of
networks with different mixing patterns [192], or epidemic spreading [75,77]. As in the interdependent network case, the
percolation threshold depends on the different types of degree correlations between the layers. Ref. [48] investigated and
discussed also the role of targeted attacks of multiplex networks. Finally, in Ref. [193] the percolation transition has been
characterized for a network in which each layer is a subgraph of some underlying network.
4.6.2. Percolation of antagonistic networks

In Ref. [194], the Authors considered the case in which there is antagonism between the different layers of a multiplex.
In particular, they studied the case of a duplex, and they assumed that the function or activity of a node is incompatible with
the function or activity of the replica node in the other layer. The nodes in layer α are indicated in the following by (i, α),
while the nodes in layer β are indicated by (i, β), with i = 1, 2, . . . , N. The two nodes (i, α) and (i, β) are replica nodes
with an antagonistic interaction. In order to determine if a node (i, α) is in the percolation cluster of the α , or if a node (i, β)
is in the percolation cluster of the layer β , the following algorithm has been proposed.
A node (i, α) is part of the percolation cluster in layer α if the following recursive set of conditions is satisfied:
(a) at least one neighbor (j, α) of node (i, α) in layer α is in the percolation cluster of layer α ;
(b) none of the nodes (j, β), neighbors of the replica node (i, β), are in the percolation cluster of layer β .
Similarly, a node (i, β) is part of the percolation cluster in layer β if the following recursive set of conditions is satisfied:
(a) at least one neighbor (j, β) of node (i, β) in layer β is in the percolation cluster of layer β ;
(b) none of the nodes (j, α), neighbors of the replica node (i, α), are in the percolation cluster of layer α .
For a given multilayer network which is locally tree-like it is possible to construct a ‘‘message passing’’ algorithm that
determines if node (i, α) belongs to the percolation cluster of layer α .
We denote by σiα→jα = 1, 0 the message within a layer, from node (i, α) to node (j, α). σiα→jα = 1 indicates that node
(i, α) is in the percolation cluster of layer α when we consider the cavity graph by removing the link (i, j) in network α .
In addition, we indicate with siα = 0 a node that is removed from the network as an effect of the damage inflicted to the
network, otherwise siα = 1.
The message passing equations take the following form:
 
 
σiα→jα = siα 1 − (1 − σℓα→iα ) (1 − σℓβ→iβ ),
ℓ∈Nα (i)\j ℓ∈Nβ (i)
5
4.5
4
3.5 II-B III
3
ZB
2.5
2
1.5
1
0.5 I II-A
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
ZA
Fig. 30. Phase diagram of the percolation of two antagonistic Poisson networks with average degree ⟨k⟩A = zA and ⟨k⟩B = zB for p = 1. In region I none of
the two networks is percolating. In Region II-A network A is percolating but not network B. Symmetrically, in region II-B network B is percolating but not
network A. In region III there are two bistable solutions to the percolation problem and, depending on the initial conditions, either network A or network
B is percolating.
Source: Reprinted figure with permission from Ref. [194]. © IOP Publishing Ltd and Sissa Medialab srl. Reproduced by permission of IOP Publishing. All
rights reserved.
 
 
σiβ→jβ = siβ 1 − (1 − σℓβ→iβ ) (1 − σℓα→iα ). (175)
ℓ∈Nβ (i)\j ℓ∈Nα (i)
Siα = 1 (Siα = 0) indicates that the node (i, α) is (is not) in the percolation cluster of layer α . The variables Siα can be
expressed in terms of the messages, i.e.
 
 
Siα = siα 1− (1 − σℓα→iα ) (1 − σℓβ→iβ ),
ℓ∈Nα (i) ℓ∈Nβ (i)
 
 
Siβ = siβ 1 − (1 − σℓβ→iβ ) (1 − σℓα→iα ). (176)
ℓ∈Nβ (i) ℓ∈Nα (i)
If we consider a duplex formed by two random networks with degree distribution PA (k) and PB (k) respectively, we can
average the messages in each layer getting the equations for σA = ⟨σiα→jα ⟩ and σB = ⟨σiβ→jβ ⟩, that read as
σA = p[1 − GA1 (1 − σA )]GB0 (1 − σB )

σB = p[1 − GB1 (1 − σB )]GA0 (1 − σA ). (177)
The probabilities SA , (or SB ) that a random node in layer A (or layer B) is in the percolation cluster of layer A (or layer B) are
given by
SA = p[1 − GA0 (1 − σA )]GB0 (1 − σB )
SB = p[1 − GB0 (1 − σB )]GA0 (1 − σA ). (178)

This novel percolation problem has surprising features [194]. For example, for a duplex network formed by two Poisson
networks with average degree respectively ⟨k⟩A = zA and ⟨k⟩B = zB , the phase diagram is shown in Fig. 30, and displays a
bistability of the solutions for some parameter values. In region I, for zA < 1 and zB < 1 none of the two Poisson networks
are percolating. In Region II-A only network A contains a percolation cluster. Symmetrically, in region II-B only network
B contains a percolation cluster. Nevertheless, in region III, the solution of the model is bistable, and depending on the
initial condition of the message passing algorithm, either network A or network B is percolating. Therefore, this percolation
problem can display an hysteresis loop, as shown in Ref. [194].
4.6.3. K-core percolation

In Ref. [195], the k-core of multiplex networks was defined as the set of nodes between which there are at least
kα ≤ 2 paths connecting them in each layer α = 1, 2, . . . , M. It has been shown that the k-core on a duplex emerges
always discontinuously and for every degree distribution as a function of the fraction 1 − p of removed nodes, following
a hybrid transition, with the only exception of the trivial (1, 1)-core. The characterization of the k-core composition of
a multiplex networks is a significantly richer problem than the characterization of the k-core structure of single-layer
networks. Moreover, it has been shown that multiplex networks are also less robust than their counterpart single-layer
ones, if one considers the destruction of the k-cores induced by random removal of the nodes.
4.6.4. Viability in multiplex networks and weak percolation

Different models have been recently proposed that generalize the concept of percolation to interdependent networks.
Ref. [196] assumed that each layer has a certain number of resource nodes. Each node of the multiplex will then belong to
the viable cluster of the resource network, if it is connected by at least one path to a resource node in each layer. The model
displays discontinuity, bistability, and hysteresis in the fraction of viable nodes with respect to the density of networks
and the fraction of resource nodes. The model can be used in the context of infrastructures, to show that the recovery from
damage can also be extremely slow and difficult, due to hysteresis. In Ref. [197] the weak bootstrap percolation and the weak
pruning percolation have been proposed. The weak bootstrap percolation describes the recovery of multiplex networks,
while the weak pruning percolation describes the deactivation/damage propagating processes in a multiplex structure.
While bootstrap percolation and pruning percolation on single-layers are equivalent, Ref. [197] shows that they do differ in
the multiplex case. When layers are formed by Poisson networks, the weak bootstrap percolation includes both continuous
and discontinuous transitions, while the weak pruning percolation has only a continuous transition when defined on a
duplex, and a discontinuous hybrid transition when defined on a multiplex with 3 or more layers.
4.7. Cascades on multilayer networks
In multilayer networks, cascades can be caused not only by interdependent percolation but also by other processes
including cascades of loads [81,198] described, for instance, by the Bak–Tang–Wiesenfeld sandpile model [199], by the
threshold cascade model [50] proposed by Watts in Ref. [200], or by novel cascade processes [201]. In Ref. [81] cascades
of loads have been proposed to study the robustness of multilayer infrastructures and power-grids. In that work, it has been
shown that moderately increasing the interlayer connectivity can reduce the size of the largest cascade in each layer, and it
is therefore advantageous. Nevertheless, it has been shown that too much interconnections can become detrimental for the
stability of the multilayer system because it allows single layers to inflict lager cascades to neighbor layers, and furthermore
new interconnections increase the capacity and the total possible load, eventually generating larger cascades.
In Ref. [198] the role of multiplexity in the sandpile dynamics was further explored, showing that the avalanche size dis-
tribution is not affected by the multiplexity of the network. However, the detailed dynamics is affected by the multiplexity.
For example, high-degree nodes in multiplex networks fail more often than in single-layer networks. In Ref. [50] cascades
of adoption of a behavior, such as ideas or financial strategies, or movement to join, are studied with the Watts model [200].
It is shown that multiplexity can actually increase the network vulnerability to global cascades. This has relevance for ad-
vertising strategies that can become more effective with an increasing multiplexity, or for banks that might become more
vulnerable with every new layer in the multiplex. In Ref. [201] a novel cascade model, inspired by the Watts threshold model,
has been considered in which some nodes adopt a behavior once enough neighbors in all the layers have also adopted it, and
some other nodes adopt the behavior only if enough neighbors in some layers do. This model can either inhibit or facilitate
the cascade of events. Moreover, the model in some region of the phase space, allows for abrupt emergence of cascading
events, revealing some novel mechanisms for the onset of avalanches that can be useful to characterize civil movements, or
product adoption.
5. Spreading processes and games
As in the case of single-layer networks, the characterization of multilayer networks has to be completed with the
knowledge about how the structural features of these networks affect their functioning. The ubiquity of multilayer networks
in natural, social and technological systems demands to revisit most of the dynamical settings that were previously
successfully addressed. This is the goal of this section. The most relevant processes that have been studied include stochastic
processes such as random walks, epidemic spreading and evolutionary game dynamics. In addition, each layer can sustain
different coevolving processes, that are influencing each other. Thus, the multilayer formalism allows to study the important
scenario in which different dynamical processes interplay, such as the competition between different virus strains in a
population.
The section is structured as follows. The first part is devoted to diffusion processes starting from the simple cases of
linear diffusion and random walks. As a generalization of these processes, we also address the onset of congestion in
communication systems. In this latter framework, the diffusion in the network takes into account the limited capacity of
the nodes for sending and storing incoming packets. There, the interest lies on uncovering how the multiple routes between
network layers affects the onset of jamming in the system.
The second part of the section focuses on spreading processes. In this case, the main goal is to revisit the methods
developed for epidemic spreading on complex networks and generalize them to the framework of multilayer. As usual in this
context, the main concern is the prediction of the outbreak threshold. However, other measures such as the effect of potential
control measures, the coevolution of the disease spreading, and the transmission of information about its impact can be
addressed by coupling the dynamical processes in each of the network layers. In addition to disease spreading, another type
of contact process, i.e. opinion dynamics, will be reviewed. In this setting, the goal is identifying the conditions for reaching
agreement or consensus that, in the context of multilayer networks, can be extended to a problem for the coevolution of
different opinions in a system of agents.
Finally, the third part is concerned with the emergence of cooperative traits, and their survival under the framework
of natural selection. This topic is quite challenging since defective behaviors are seen to provide more benefits at an
individual level, leading to the extinction of cooperation. In the last decade, spatial evolutionary games have been extensively
investigated showing that cooperation is promoted thanks to the possibility of forming compact cooperative clusters.
This mechanism coined network reciprocity turned out to be enhanced in single heterogeneous networks. One interesting
question naturally arises: could network reciprocity be further enhanced? Many different ways of coupling the evolutionary
dynamics in each network layer have been proposed, and we will here systematically review the main achievements in this
direction.
5.1. Diffusion on multilayer networks
One of the most general issues on networks is the study of transport processes across network topologies [202,203].
These studies, indeed, are good proxies for many real dynamics involving navigation and flows of many kinds, data in
computer networks, resources in technological ones or goods and humans in transportation systems. The variety of such
processes ranges from simple diffusive dynamics and random walks [204–208] to sophisticated protocols aimed at routing
information [209–212]. These efforts allowed to characterize many central features of networks that cannot be captured
from intrinsically local properties of nodes, and to design optimal navigation schemes suited to real network architectures.
5.1.1. Linear diffusion

The study of transport processes in multilayer networks has allowed to add an intrinsic ingredient of real transportation
and communication systems, which is the possibility of driving flows by using multiple channels, or modes. The first attempt
to address this issue was done by Gómez et al. in Ref. [45], in which simple linear diffusion is generalized to the case of
multiplex networks. The time-continuous equations for the linear diffusion in a multiplex composed of M networked layers
(each of them having one representation of each of the N nodes) are:
dxαi
N M
β
wijα (xαj − xαi ) + Dαβ (xi − xαi ),
 
= Dα (179)
dt j =1 β=1
where xαi (t ) (i = 1, . . . , N and α = 1, . . . , M) denotes the state of node i in layer α . The diffusion coefficients are, in
general, different for each of the intralayer diffusive processes (Dα with α = 1, . . . , M) and for each of the interlayer
diffusion dynamics (Dαβ with α , β = 1, . . . , M). In this way, the first term in the right-hand side of Eq. (179) accounts
for the intralayer diffusion, while the second one represents those interlayer processes. Being a multiplex, the latter ones
take place between the replicas of a given node in each of the layers.
The above dynamical equations can be casted in the usual form by defining the Laplacian L (usually called supra-
Laplacian) of the multiplex as an MN × MN matrix of the form:

...

D1β I −D12 I −D1M I
 β̸=1
D1 L1 ...
 
0 0  
...
 
 0 D2 L2 ... 0   −D21 I D2β I −D2M I 
  β̸=2 ,

L= .. .. .. .. + (180)
 
. .. .. .. ..

. . .  
 
. . . .
 
... M  
0 0 DM L 
−DM1 I −DM2 I ... DM β I
 
β̸=M
where I is the N × N identity matrix and Lα is the usual N × N Laplacian matrix of the network layer α whose elements are
Lαij = sαi δij − wijα , and sαi is here the strength of the representation of node i in layer α , sαi = j w α
ij . With this definition of
L, the diffusion Eq. (179) can be formulated as ẋ = Lx.
For the sake of simplicity, one considers that the intralayer diffusion coefficients are set to Dα = 1 ∀α (without any loss
of generality, as each coefficient Dα can be absorbed by the terms of the corresponding Laplacians Lα ) and that the interlayer
coefficients are equal to Dαβ = Dx ∀ α, β . The simplified Laplacian thus reads:
L1 ... (M − 1)I ...

 
0 0 −I −I
 
0 L2 ... 0  −I (M − 1)I ... −I
.

L=. .. .. ..  + Dx ·  .. .. .. .. (181)
 
. . . . . 

. . . .

0 0 ... LM −I −I ... (M − 1)I
Fig. 31. (Color online) Evolution of λ2 (Dx ) for multiplexes composed of two layers and N = 103 . The layers are of the form: (a) scale-free (SF) with
P (k) ∼ k−2.5 and Erdős–Rényi network with average degree ⟨k⟩ = 8, (b) same SF network as in (a) and a Watts–Strogatz small-world network with
⟨k⟩ = 8 and replacement probability of 0.3, (c) same SF network as in (a) and another SF network with P (k) ∼ k−3 and (d) same SF network as in (a) and
the same SF network plus 400 additional links randomly placed. It is clear that in the first three examples superdiffusion shows up for large values of Dx ,
whereas for the multiplex in (d) this is not the case.
The above formulation allows to study the role of the interlayer coupling Dx in the diffusion properties of the multiplex,
and relating them with those of the set of isolated network layers (Dx = 0). In fact, given the symmetry of the Laplacian
(181), the solution of the linear system can be expressed in terms of the eigenvectors and eigenvalues of L whose time
evolution reads φi (t ) = φi (0)e−λi t . Thus, the diffusion time-scale, i.e. the time τ that the multiplex needs to relax to the
stationary solution, is given by the inverse of the smallest non-zero eigenvalue of L that, for a connected multiplex, is the
second eigenvalue (from the smallest λ1 = 0, to the largest one), that is τ = 1/λ2 .
In Ref. [45] the Authors focused on the case of two layers, M = 2, and through perturbation limit they were able to relate
the value of λ2 for the weak (Dx ≪ 1) and strong (Dx ≫ 1) interlayer coupling limits. In Ref. [59], Solé et al. extended the
perturbative analysis to the full spectrum of L, {λ1 = 0, λ2 , . . . , λM ·N }, and for an arbitrary number M of layers. The main
results of Ref. [59] concerning the smallest non-trivial eigenvalues of L are the following:
• For Dx ≪ 1 the set of the M − 1 smallest non-trivial eigenvalues of L increases linearly with Dx .
• For Dx ≫ 1 the set of the N − 1 smallest non-trivial eigenvalues of L are approximately the N − 1 non-zero eigenvalues
of a Laplacian corresponding to a simple network of N nodes resulting from taking  the average of the layers composing
α
the multiplex, so that the weights of each link in this latter network read as wijav = α wij /M.
The main result of Refs. [45,59] appears evident when looking at the strong coupling regime, Dx ≫ 1. While for Dx ≪ 1
the time scale associated to diffusion goes as τ ∝ (M · Dx )−1 , the case Dx ≫ 1 shows that the multiplex is always faster than
the slowest one of the composing layers. Moreover, although not guaranteed, superdiffusion (i.e. the fact that the multiplex
relaxes faster than any of its composing layers) is also possible. In Fig. 31 we show the evolution of λ2 (Dx ) for four multiplexes
(with and without superdiffusion) composed of two layers.
From the evolution of λ2 (Dx ) shown in Fig. 31 it is clear that the diffusive properties (as described by τ ) of the multiplex
do not evolve homogeneously with the interlayer coupling, and there is a value (Dx )c (that depends on the composition of the
multiplex) in which the linear increase of λ2 (Dx ) changes abruptly. This phenomenon is explained by Radicchi and Arenas in
a e
2.0
1.5
λ2
1.0
0.5 b
0.5
<vA|vB>
0.0
-0.5 c
<vA|1>, <vB|1>
-5 d
10-1 100 101 102
Dx
-0.1 0.0 0.1 -0.1 0.0 0.1
characteristic valuation characteristic valuation
Fig. 32. (Color online) Evolution of λ2 and its associated eigenvector, v2 , as a function of Dx for a multiplex composed of two Erdős–Rényi networks of
N = 50 nodes and average degree k̄ = 5. In this example, the critical point is (Dx )c = 0.602(1). (a) Values of the components in v2 (characteristic
valuations) for the nodes in the two layers for Dx = 0.602 (just before the onset of the transition). (b) λ2 as a function of Dx (black
 line).
 The discontinuity of
λ2 LA +LB
the first derivative of λ2 is very clear. The transition between the two known different regimes (2Dx , blue dashed line, and 2
, red dash-dotted line)
is evident. (c) Projection of vB2 into vA2 as function of Dx . (d) Projection of the unit vector into vA2 and vB2 as functions of Dx . These two projections indicate
the sum of all components of vA2 and vB2 respectively. (e) Values of the components in v2 (characteristic valuations) for the nodes in the two layers for
Dx = 0.603 (just after the onset of the transition).
Source: Reprinted figure from Ref. [96]. Courtesy of A. Arenas.
Ref. [96] as a structural transition from a decoupled regime (in which layers behave as independent) to a systemic regime (in
which layers are indistinguishable). This change is also explained in terms of the eigenvector v2 associated to the smallest
non-zero eigenvalue of the Laplacian of the multiplex, λ2 , since it provides useful information about the modular structure
of a network, as it is the key ingredient of the spectral partitioning method developed by Fiedler [213–215]. In Ref. [96],
the Authors focused on multiplexes composed of 2 layers A and B and recovered, by solving a minimization problem, the
same result as in Ref. [45] for the behavior of λ2 for small and large values of Dx . Then, they analyzed the evolution with Dx
of the components of the vector v2 focusing on the difference between its first N components (those corresponding to the
first layer A), vA2 , and the remaining N (those corresponding to the second layer B), vB2 . They observed, as in Ref. [45], that
for Dx ≤ (Dx )c the components of v2 are of the form vA2 = −vB2 (see panel (a) of Fig. 32), each one of them having all the N
entries identical. On the other hand, when Dx ≫ 1, the components of each subvector have the same sign (see panel (e) of
Fig. 32). Thus, in this limit, the partition of the multiplex into two layers is no longer possible. Furthermore, they found an
upper bound for the value (Dx )c as:
1
(Dx )c ≤ λ2 (LA + LB ), (182)
4
where LA and LB are the Laplacian matrices of layers A and B, respectively. The structural phase transition showing the exact
value of (Dx )c is apparent in panel (c) of Fig. 32 when measuring the evolution of the projection of vB2 into vA2 as function
of Dx .
5.1.2. Random walks

Although linear diffusion provides already a deep understanding of the differences between the behavior of single and
multilayer networks, the connection with diffusion phenomena on real systems is not so straightforward due to the discrete
nature of flows in real communication and transport systems. In Ref. [98] De Domenico et al. generalized random walks in
monolayer graphs to the case of multiplex networks. Fig. 33 is an illustration of a path of a random walker (RW) exploring a
multiplex network composed of N = 7 nodes and M = 3 layers, in which intralayer and interlayer transitions are shown.
In Ref. [98] a general form for the evolution of the occupation probabilities of a RW in a multiplex of M layers and N nodes
is written as
M N M 
N
βα βα
pi,α (t + 1) = Πiiαα pi,α (t ) + Πjiαα pj,α (t ) +
  
Πii pi,β (t ) + Πji pj,β (t ) (183)
β=1 j=1 β=1 j=1
β̸=α j̸=i β̸=α j=
̸ i
Fig. 33. (Color online) Path (sequence of arrows) of a random walker navigating a multiplex network composed of N = 7 nodes and M = 3 layers. In this
example, the walker is neither allowed to switch between layer 1 and layer 3 in one time step nor to change node and layer simultaneously.
Source: Reprinted figure from Ref. [98]. Courtesy of A. Arenas.
where pi,α (t ) is the probability of finding the walker at time t in the representation of node i in layer α . The transition
αβ
probabilities are encoded in 5 so that Πiiαα is the probability that the walker does not move, Πii is the probability that a
αα
walker changes from layer α to β staying in the same node i, Πij is the probability that the walker hops from node i to j
αβ
within the same layer α , and Πij is the probability that the walker changes from node i to node j, and also from layer α to
β (this last movement being not actually represented in Fig. 33).
By setting the values of each transition probability, the Authors of Ref. [98] studied different kinds of RWs. As in the case
of single-layer networks, the mathematical formulation of RWs makes use of the adjacency matrix, A, of the multiplex (also
called supra-adjacency matrix) in order to write the expressions for the transition probabilities between nodes and layers.
The matrix is of the form
W1 ...
 
D12 I D1M I
 D21 I W2 ... D2M I
A= . .. ..  , (184)
 
.  . . . 
DM1 I DM2 I ... WM
where Wα is the adjacency matrix of layer α with elements {wijα }. Then, the generalization of the classical RW to a multiplex
is obtained by setting
Πiiαα = (1 − q), (185)
αβ Dαβ
Πii = q , (186)
sαi + Siα
αβ wijα
Πii = q , (187)
si + Siα
α
αβ
Πii = 0. (188)
α

In the above expressions Si = β=1 MDαβ is the strength of the interlayer coupling of the representation of node i in layer
α , and q is the hopping rate such that 1 − q accounts for the possibility of the random walker to wait in a given vertex and
in a certain layer. In fact, considering the typical assumption in classical RWs, one can set q = 1 so that Πiiαα = 0 (i.e. the
walkers never remain in the same position) and, following again the simplification done for linear diffusion, one can further
assume that Dαβ = Dx ∀α , β .
With the help of these assumptions, one can study the diffusion properties of RWs. In Ref. [98], the coverage of the walks
(measuring the number of nodes visited at least once in a given time window τ ) is studied in detail for multiplexes composed
of two layers. The main result is that for large enough values of Dx , the time to cover the multiplex is smaller than the whole
required to explore each one of the layers separately.
Another important measure (when dealing with RWs) is the concept of communicability [216,217], that quantifies the
number of possible paths communicating two given nodes of a network. In the same framework of classical RWs in multiplex,
communicability has also been extended in Ref. [218] to account for the all the walks between any pair of nodes in a
multiplex. Indeed, considering the supra-adjacency matrix of Eq. (184), the number of walks of length L between the
representation of a node i in layer α and the representation of node j in layer β is given by different entries of the Lth
power of the supra-adjacency matrix, AL . Considering the definition of communicability in single-layer networks, one is
interested in assigning more importance to shorter walks than to long ones. In this way, one defines the communicability
matrix G as
∞
A2  AL
G=I+A+ + ··· = = exp(A) . (189)
2! L=0
L!
One can further express the communicability matrix as
G1 G12 ... G1M

 
 G21 G2 ... G2M 
G = exp(A) =  . .. ..  , (190)
 
.  . . . 
GM1 GM2 ... GM
where Gα is a N × N matrix containing the communicability between the representations of every pair of nodes within
layer α of the multiplex. It is important to note that Gα ̸= exp(Wα ), since Gα takes into account those paths that, while
connecting the representations of two nodes within the same layer, can include hops to any other layer different from α due
to the coupling Dx . Obviously, if Dx = 0, then one has
exp(W1 ) ...
 
0 0
 0 exp(W2 ) ... 0 
G= .. .. .. . (191)
 
 . . . 
0 0 ... exp(WM )
This means that communicability naturally accounts for the fact that, in a multiplex, information between the representation
of two nodes within the same layer flows actually in parallel at the different layers.
5.1.3. Diffusion-based centrality measures

Another relevant aspect of diffusive processes is the evaluation of the importance that the nodes of a network have
on its global organization. To this aim, RWs have intensively been used to assess the importance of nodes based on their
occupation number, i.e. the probability of finding a walker in each of the nodes in the steady state of the dynamics. In Ref. [98],
the Authors study the set of occupation numbers in the steady state of the dynamics, {p∗i,α }, for the general family of walks
described in Eq. (183), including the classical one, by assigning different values for the transition probabilities included in 5.
In Ref. [125], the Authors focused on the generalization of the classical RWs (with q = 1). Since the occupation probabilities
in a multiplex refer to each representation of a given node i on a given layer α , p∗i,α , in order to evaluate the centrality of a
given node i one must aggregate the occupation numbers for each of its representations in all the layers
M

p∗i = p∗i,α , (192)
α=1
so that p∗i accounts for the probability of finding the walker in node i, regardless of the layer.
One of the most successful applications of RWs to derive centrality measures is the well-known PageRank centrality
method [84,219], which is at the core of the most widely-used search engine of the World-Wide-Web, Google. PageRank
is based on a diffusion process that mimics the navigation through webpages as the motion of a random walker following
hyperlink pathways. Whereas these link-by-link hops are followed most of the times, there is a small probability r that at
a given time step the walker performs a long distance hop to an arbitrary node of the system. In this case, the transitions
encoded in 5 should be reformulated in the following way 5̂ = r 5 + (1 − r )H, with r ∈ [0, 1], where (1 − r ) is the
αβ
probability of performing a long distance hop and the elements of H are defined as Hij = 1/(MN ).
In Ref. [54], Halu et al. proposed a different version of the PageRank dynamics that is not as general as the above in the
sense that there is a privileged layer that influence the remaining layers of the multiplex. In the case of a 2-layer multiplex,
the nodes in the first one are characterized by means of the usual PageRank on single-layer networks, whereas the transitions
between adjacent nodes of the second layer are biased by the centrality of their counterparts in the first layer.
Finally, in Ref. [125] the Authors use the formalism provided by RWs in a multiplex to compute other centrality measures
not directly computed from the occupation numbers of the nodes, but with the number of paths that cross them. Examples
are betweenness and closeness centralities, which in principle are designed to account for shortest paths, but that can be
extended to RW paths.
5.1.4. Routing and congestion phenomena

The processes described in this section are proxies of real diffusion dynamics in transportation and information
interconnected networks. In fact, most of the data and human flows in such systems are far from random, and most of the
times they follow paths that minimize a distance or certain other cost functions. This is the case of shortest-path diffusion
through which the agents (such as data packets or humans) go from an origin node to a destination one following the path
that minimizes the number of nodes visited along their trip (in the case of unweighted networks), or the sum of the inverse
of the weights associated to each link contained in it (for weighted ones). In this context, the problem of congestion arises
when the so-called capacity of nodes is limited, i.e. the number of packets that a node can send to its neighbors per unit
time is bounded. However, in a time step, it may occur that a node receives more packets than its capacity. When this event
happens for a macroscopic part of the system, congestion shows up.
The usual set up for studying congestion is a situation in which a number of packets, pN, is injected in the network per unit
time, being the origin and destination nodes of these packets randomly assigned, and p the probability of an information
packet is created. Packets are delivered from the origin to their destination (where they disappear) by hopping between
adjacent nodes. As congestion results in the unbalance between the number of packets introduced in the network and that
of delivered ones, the following order parameter [220,221]
A(t + T ) − A(t )
ρ = lim (193)
t →∞ pNT
is used to monitor the transition to congestion. In the latter equation, A(t ) is the number of packets that the network is
handling at some time t, and thus ρ measures the normalized (0 ≥ ρ ≥ 1) increase of packets per unit time. Many works
in single-layer networks (see for instance Refs. [203,211,220–223]) have studied the transition from the free-flow regime
(ρ = 0) to the congested one (ρ > 0) at some critical load of the network pc . Many different ways of delivering the packets
on top of a variety of network topologies have been introduced in order to delay the onset of congestion.
Recently, several works have addressed the problem of congestion [92,224–226] on top of interconnected networks.
In Ref. [92], Morris and Barthélemy analyzed an interconnected system formed by two layers which partially resembles a
multiplex. In fact, the second layer is composed of a subset of nodes of the first one so that each node in the second layer
is directly connected with its representation in the first network. Origin and destinations can be only chosen in the first
(large) layer so that, given that the agents travel following the shortest path, when networks are uncoupled the agents do
not make use of the second platform. However, coupling enters into play when the diffusion of agents between adjacent
nodes is faster in the second of the layers. In this case, agents may find short paths between their origin and destination
making use of the second layer. To account for this coupling, the Authors measure the coupling between layers as
 σijcoupled
λ= Tij , (194)
i̸=j
σij
coupled
where Tij is the fraction of agents traveling from node i to node j (of the first layer) whereas σij and σij are, respectively,
the number of shortest paths between the former two nodes i and j, and those paths that contain edges from both network
layers. Obviously λ ∈ [0, 1], being λ small (large) in the weakly (strongly) coupled regime. In fact, as λ increases, the average
distance between two nodes in the first layer decreases as an effect of the use of the second layer. However, this effect
comes together with the accumulation of the used paths on a very small subset of edges which can lead to a saturation of
the system due to congestion. This latter effect is shown by measuring the Gene coefficient G ∈ [0, 1] that accounts for the
distribution of paths across the edges, so that small (large) values of G correspond to a spread (concentrated) distribution of
the flows across the edges (see Fig. 34). The Authors show that it is possible to obtain an optimal coupling λ⋆ that balances
the aforementioned effect, so to obtain an optimal communication multiplex.
The congestion diagram ρ(p) has been studied for usual routing protocols implemented on top of interconnected
networks in Refs. [224–226]. As described above, congestion can be directly connected to the limited capacity of nodes
and links for handling the flows. In these latter works, the Authors consider interconnected networks in which some pairs
of nodes of different networks are linked, thus providing the communication between the two networks and the mix of
their flows. The main conclusion [224,226] is that the more is the interconnection between networks, the better is the
performance of the system in terms of the critical value pc from which the system gets congested. Moreover, in Ref. [226] the
Authors show that when nodes of similar degrees from different networks are linked, the critical value pc raises considerably,
pointing out that the connection of the hubs of the different networks helps to alleviate the saturation of both networks.
5.2. Spreading processes
Among the possible dynamical scenarios of relevance, the analysis of the spreading of diseases and the design of
contention policies have capitalized most of the literature in the field of network science [4,8,9,227–229]. This interest
has been rooted in the successful extension of the classical compartmental models used in mathematical epidemiology
[230–233] beyond the mean-field equations, so to incorporate networked architectures that capture realistic interaction
patterns. From the physical point of view, the study of epidemic models on top of networks has allowed to extend a paradigm
Fig. 34. (Color online) Evolution of the average Gene coefficient ⟨G⟩ of the distribution of paths across the edges (left panel) and the average distance
⟨d̄⟩ between nodes (right panel) as functions of the average value of the interlayer coupling ⟨λ⟩ in the model introduced in Ref. [92]. The different curves
correspond to different organization of the origin–destination matrix Tij . This organization is initially centralized (one node is the unique destination of all
the packets) and then randomized with some probability p. Reported data correspond to: p = 0 (purple dots), 0.2 (blue squares), 0.4 (green diamonds), 0.6
(orange triangles), and 0.8 (red inverted triangles).
of non-equilibrium phase transitions beyond lattice models [234]. In addition, the study of the spreading of infectious
diseases on top of networks gives also insight on other dynamical processes of interest, such as the information and rumor
spreading in social networks. Thus, the recent interest on multilayer networks has provided a natural way to extend and
improve our understanding of the basics of spreading dynamics in real-world complex systems.
Let us first briefly review two typical compartmental models: the Susceptible–Infected–Susceptible (SIS) and the
Susceptible–Infected–Recovered (SIR) models. In both cases susceptible individuals can be infected when they are in contact
with an already infected agent. In this case, and considering the time discrete version of SIS and SIR dynamics, there is some
probability β that the individual takes the infection. Once infected, agents recover with some probability µ per unit time.
The difference between SIS and SIR models is the following: in the SIS model recovered individuals become susceptible of
contracting the disease, i.e. there is no acquired immunity after suffering the infection (this setting suits with most of the
typical sexually transmitted diseases). At the opposite, in the SIR case, recovered individuals are immune to the disease,
and they cannot be infected again. From the dynamical point of view, SIS and SIR models differ in their stationary regime:
while for the SIR model the dynamics dies out once there is no infected individual in the population (so that only susceptible
and recovered agents coexist), for the SIS model sustained epidemic activity may persist over time as a consequence of the
reinfections of recovered agents.
In any case, the key question is whether the spread of the disease causes an epidemic scenario in which a macroscopic
part of the population is affected. The typical control parameter in both SIR and SIS models is the ratio between the infection
and recovery probabilities, β/µ, while the order parameter that quantifies the impact of the disease depends on the type of
dynamics used. In the case of the SIR model, the impact of the disease is given by the number of recovered individuals in the
final state of the dynamics. For the SIS model, instead, the order parameter is the number of infected agents in the sustained
endemic state, that allows to track a non-equilibrium transition between the healthy and the epidemic phases.
5.2.1. Disease spreading on single-layer networks

We start by briefly summarizing the basic equations that capture a disease spreading process on top of a single-layer
network. To this aim, we will focus on the SIS model, although other compartmental dynamics such as the SIR model can be
casted under the same framework, and we will review the microscopic formalism developed in Ref. [235].
As usual, we consider a network of N nodes that, in the most general case, is weighted. We represent the connections
between the nodes by the entries {wij } of a N × N weighted adjacency matrix W. Under the SIS framework, at each time step,
an infected node makes a number λ of trials to transmit the disease to its neighbors with probability β per unit time. Since
this process has no memory, the probability pi (t ) that node i is infected at time t follows a Markovian dynamics defined as
pi (t + 1) = (1 − qi (t ))(1 − pi (t )) + (1 − µ)pi (t ), (195)
where qi (t ) is the probability that node i is not being infected by any neighbor at time t,
N

qi (t ) = (1 − β rji pj (t )), (196)
j =1
and rij are the elements of a N × N matrix R representing the probability that node i is in contact with node j, that is
λ
wij

rij = 1 − 1− . (197)
si
The first term on the right hand side of Eq. (195) is the probability that node i is susceptible (1 − pi (t )) and is infected
(1 − qi (t )) by at least one of its neighbors, whereas the second one is the probability that node i is infected at time t and does
not recover. The form of matrix R in Eq. (197) allows to capture different kinds of interactions under the same mathematical
formulation. For instance, the case λ = 1 (one contact between i and j per unit time) corresponds to a Contact Process (CP),
so that rij = wij /si , being si the total strength of node i. Another general situation is that corresponding to λ → ∞, which
yields rij = aij , leading to a Reactive Process (RP), regardless of whether the network is weighted or not. The latter limit is
what is typically adopted for the SIS dynamics.
The phase diagram of the SIS model is then obtained by searching for the stationary solution {p∗i } of Eq. (195), that results
in the following equation
p∗i = (1 − q∗i ) + (1 − µ)p∗i q∗i , (198)
and computing the usual order parameter of the SIS model as
N
1  ∗
ρ= pi . (199)
N i =1
Eq. (198) has always the trivial solution pi = 0, ∀i = 1, . . . , N (corresponding to ρ = 0). However, other solutions
corresponding to ρ > 0 are possible, and they represent the existence of an epidemic state. The relevant question is thus
what is the critical value of the control parameter β/µ for which ρ > 0 is a solution of the SIS dynamics.
The calculation of this critical point is performed by considering that, when the value β/µ is close to the critical one the
probabilities pi ≈ ϵi (0 < ϵi ≪ 1). The substitution in Eqs. (196) and (198) yields
N

qi ≈ 1 − β rji ϵj , (200)
j =1
N 
µ
 
0= rji − δji ϵj , (201)
j =1
β
where δij = 1 for i = j, and δij = 0 otherwise. The system of Eq. (201) shows that a non-trivial solution is possible when µ/β
is an eigenvalue of the matrix R. Thus, the problem of finding the epidemic threshold (the minimum value of β/µ giving
rise to an epidemic) is tantamount to computing the spectra of the contact matrix R. In particular, since we are interested
in finding the smallest as possible value of β/µ, we are interested in the largest eigenvalue, Λ = max(λ1 , λ2 , . . . , λN ), of
the matrix R. For a given value of µ, the critical value for the probability of infection, βc , that represents the onset of the
epidemic phase, is
µ
βc = . (202)
Λ(R)
It is worth analyzing the two limiting cases of CP and RP. In the first case, one obtains the trivial result that the only
non-zero solution corresponds to βc = µ, because the matrix R is a transition matrix whose maximum eigenvalue is always
Λ = 1. For the RP corresponding to the SIS spreading process, instead, the largest eigenvalue must be calculated [236].
We observe here that the determination of the epidemic threshold for single-layer networks is a subject of active debate
[228,229]. Indeed, there is one potential problem in the derivation of Eq. (202) coming from the fact that the maximal
eigenvector of the matrix rij might be localized on an infinitesimal fraction of nodes in the network. In this case, the order
parameter would be non vanishing only on an infinitesimal fraction of the network and, therefore, it is not possible to speak
of a true phase transition at βc determined by Eq. (202). Finally, for annealed networks which rewire their connections
on the same time scale of the dynamical process and keep their degree sequence constant, one can approximate rij as
kk
rij = ⟨aij ⟩ = ⟨ik⟩j . This matrix has maximal eigenvalue Λ′ = ⟨k2 ⟩/⟨k⟩, recovering for these networks [237,238] the classical
result reported in Ref. [239].
The above result was originally obtained by means of the so-called heterogeneous mean-field (HMF) approximation
developed in Refs. [239–241]. The HMF assumes that the contact matrix rij can be approximated as rij = ⟨aij ⟩, and
considers that all nodes having the same degree behave equally which, in terms of the microscopic formulation, means
that pi (t ) = pj (t ) whenever ki = kj . Under this assumption, valid if the network is dynamically rewiring its connections
keeping the degree distribution fixed, one ends up with a system of equations for each degree class k present in the network.
kk
By assuming ρk (t ) to be the probability that a node of degree k is infected and assuming rij = ⟨aij ⟩ = ⟨ik⟩j , the HMF
equations read

ρk (t + 1) = ρk (t ) − µρk (t ) + kβ(1 − ρk (t )) P (k′ |k)ρk′ (t ), (203)
k′
where P (k′ |k) is the conditional probability that, given a node of degree k, this is connected with another one
with degree
k′ . The stationary solution, {ρk∗ }, of the above set of equations gives the value of the order parameter ρ = N k P (k)ρk∗ . In
analogy with the microscopic formulation, one is interested in the non-trivial solutions (yielding ρ > 0) of

µρk∗ = kβ(1 − ρk∗ ) P (k′ |k)ρk∗′ . (204)
k′
In Refs. [239–241] it is shown that the above equation has non-trivial solutions when β/µ > ⟨k⟩/⟨k2 ⟩ (in agreement with
the result obtained via the microscopic formulation). It is important to note that this result points to the fact that for degree-
heterogeneous networks (such as SF networks) the epidemic threshold can be extremely small, since the second moment
of the degree distribution P (k) becomes very large. On its turn, the evidence of an almost absent epidemic threshold in real
systems defied the classical results on epidemiology, and demanded for a change in the usual contention policies applied in
systems in which an homogeneous structure was assumed.
5.2.2. Disease spreading on interconnected networks

In this context, two main kinds of multilayer networks have capitalized the attention in the recent years: interconnected,
and multiplex graphs. As far as disease spreading dynamics is concerned, both substrates have their own relevance, and the
generalization of compartmental models such as the SIS and SIR to these substrates has stimulated a great interest. The
case of interconnected networks considers the spreading of a disease in several network layers, each node of each network
representing a different agent. One example is the case of sexually transmitted diseases. Just think of three networks of
sexual contacts: one composed of agents sharing heterosexual relationships, and two other graphs of homosexual (gay and
lesbian) contacts. These three networks are connected through edges in which one (or both) of the connected nodes is
a bisexual agent. Obviously, the values of the parameters describing the spreading dynamics, such as the infection rates,
depend in principle on the nature of the link through which the contagion is produced. In this setting, the key question is to
relate the epidemic onset of the whole system with those corresponding to the isolated networks.
Different works have addressed the above issue in both the SIS [53,242–244] and SIR [76,245] frameworks. For the SIS
case, different approaches point out to a same central result: An epidemic may appear in the interconnected system even for
infection rates for which a disease would be unable to propagate in each isolated network. This result is obtained both via the
HMF approximation [242], and via the analysis of the microscopic equations and the spectral properties of the associated
connectivity matrix [53,243,244].
As pointed above, the SIS dynamics considers different parameters depending on the nature of the contacts and the
agents. Focusing on the case of two populations A and B, β aa (β bb ) is the infectious rate between individuals belonging to the
same network A (B) whereas β ab (β ba ) is the infectious rate from a node in network A (B) to another one in B (A). The recovery
rates are also fixed to µa (µb ) for individuals in population A (B). In this way, one can write the microscopic equations based
on the adjacency matrix of the whole interconnected system, which in the case of two coupled networks read as
   
A1 0 0 C12
Â = A + C = + (205)
0 A2 C21 0
where submatrices A1 (A2 ) define the connections between nodes of network A (B), while C12 (C21 ) define, on their turn,
the connections from nodes of network A (B) to those in B (A). In Ref. [53], the exact microscopic equations for generic SIS
dynamics are written, extending Eq. (195), which in the case of the usual Reactive scenario (and for two populations) yields
⃗(t + 1) = (1⃗ − p⃗(t )) ∗ (1⃗ − q⃗(t )) + (1⃗ − µ)

p ⃗ ∗ p⃗(t ). (206)
Note that µ ⃗ is a vector whose components can take values µ and µ depending on the network in which the corresponding
a b
agent is placed. In addition, the expression for the probability that a node i is not infected by any neighbor (in network A or
B), qi (t ), reads as

qi (t ) = (1 − Rij pj (t )), (207)
j
where the reaction matrix R of the system incorporates the infection rates into the adjacency matrix of the interconnected
networks:
β aa A1 β ab C12
 
R= . (208)
β ba C21 β bb A2
β
Fig. 35. (Color online) Panel (a): Fraction of infected individuals (ρ ) at the steady state as a function of the control parameter µ for two interconnected
networks of N = 104 nodes each for different values of the ratio η = βδ . The arrows show the inverse of the largest eigenvalues of the two coupled
networks. The inset shows the case when the networks are uncoupled. In panel (b) the fraction of infected individuals is shown for each of the networks
and for η = 2.0. The inset is a zoom around the critical point showing how the epidemic in the second network is triggered by the onset in the first one.
As in the case of single-layer networks, the spectral properties of this reaction matrix yield the critical behavior of the
system, i.e. the epidemic onset of the interconnected system. As a consequence, one can study different combinations of the
transmission parameters to assess the conditions for the emergence of an epidemic as in Ref. [243].
A useful insight about the behavior of the epidemic onset is achieved by treating equally the propagation properties of
the disease within each layer (β aa = β bb = β and µa = µb = µ), and by symmetrizing the propagation between them
β ab = β ba = δ . Consequently, one can derive interesting results based on perturbation theory [53] and spectral graph
theory [243,244]. For instance, the works reported in Refs. [53,244] show that the epidemic threshold βc is given by the
largest eigenvalue of a matrix defined as A = A + βδ C, so that
µ
βc = . (209)
Λ(A)
In fact, since the spectral radius of a submatrix is always smaller or equal than that of the whole matrix, one obtains that
Λ(A) ≥ max[Λ(A1 ), Λ(A2 )], (210)
i.e. the epidemic threshold of the whole system is always smaller than those of the isolated networks (see Fig. 35). Moreover,
in Ref. [244] the Authors analyzed the behavior of the largest eigenvalue Λ(A) as a function of the composing networks, and
concluded that whenever two nodes from A and B with large eigenvector components (corresponding to Λ(A1 ) and Λ(A2 ),
respectively) are linked, the epidemic threshold of the entire system reduces dramatically.
The fact that the interconnected system is always more prone to the epidemic than the set of isolated networks can be
explained in terms of a perturbative analysis that assumes that δ ≪ β . In Ref. [53], the Authors used the former limit to
explore the behavior when either (i) there is one dominant layer (e.g. Λ(A1 ) ≫ Λ(A2 )) and (ii) both layers have roughly
the same epidemic onset (Λ(A1 ) ∼ Λ(A2 )). In the first case, they show that the maximum eigenvalue of the whole system
is equal to that corresponding to the dominant layer Λ(A) = Λ(A1 ), whereas the nontrivial stationary solution at βc yields
⃗∗1
 
p
△ v⃗ =  δ ∗  , (211)
⃗
p
βΛ 1
where p ⃗∗1 is the stationary solution in the dominant network at its epidemic onset. It is clear that the onset of an epidemic
in the dominant layer automatically induces the spreading in the other one.
The second limit (Λ(A1 ) ∼ Λ(A2 )) reveals the strong synergetic effect of the coupling between layers. If isolated,
the networks would have a roughly simultaneous epidemic onset located at their corresponding βc . However, the
interconnection of layers makes a correction of the maximum eigenvalue as
δ p⃗∗1 C12 p⃗∗2 + p⃗∗2 C21 p⃗∗1

Λ(A) = max[Λ(A1 ), Λ(A2 )] + (212)
β (⃗p∗1 )T p⃗∗1 + (⃗p∗2 )T p⃗∗2
thus pointing out that the critical point of the system is smaller than any of the isolated layers, and that the correction to
its value depends on the product of eigenvector centralities between connected nodes belonging to different networks, as
pointed out in Ref. [244].
The HMF approximation, although being a coarse-grained picture, also recovers the above result. In this case, one can
easily generalize Eq. (203) to the case of interconnected graphs. Once again, in the case of two interconnected graphs A and
B, the HMF equations in their time-continuous version read [242],
dρkAa  
= −µA ρkAa + β aa (1 − ρkAa )kaa ρkA′ PAA (k′a |ka ) + β ba (1 − ρkAa )kab ρkB′ PAB (k′b |ka ), (213)
dt a b
k′a k′b
for the evolution of the fraction of infected individuals in network A having kaa neighbors in network A and kab neighbors in
network B (the evolution for ρkBb is analogous to Eq. (213)). Here the degree classes in each network, say A, is characterized by
a vector degree ka ≡ (kaa , kab ), and the conditional probabilities used in Eq. (213) make use of this vectorial classification of
nodes. As in the case of the HMF equations in single-layer networks, in order to derive the epidemic onset, one is interested
in the limit where the densities of infected individuals are small, and therefore one works with the linear equation
dρ
⃗
ρ + Cρ⃗ ,
= −⃗ (214)
dt
where ρ
⃗ ≡ (ρkAa , ρkBb ), and
β aa kaa PAA (k′a |ka ) β ba kab PAB (k′b |ka )

 
C= . (215)
β ab kba PBA (k′a |kb ) β bb kbb PBB (k′b |kb )
Obviously, whenever the maximum eigenvalue of matrix C satisfies Λ(C) < 1, any initial fraction of infected individuals
decays to the trivial solution ρ ⃗ As a result, the critical epidemic point is defined by Λ(C) = 1. The analysis of this
⃗ = 0.
condition in uncorrelated networks was carried out in Ref. [242], yielding that Λ(C) ≥ max[Λ(A), Λ(B), Λ(AB)], being λ(A)
and Λ(B) the maximum eigenvalues of the isolated networks, and Λ(AB) the one corresponding to that of the bipartite
network obtained by considering only the links between nodes belonging to different networks. Once again, the HMF
approximation allows to conclude that the coupled system can sustain an epidemic even when the isolated networks would
be free of disease.
For the SIR model the treatment is quite similar [76,245]. However, in Ref. [245], based on the analogy between bond
percolation and the SIR model, the Authors unveiled one important difference with respect to the case of the SIS dynamics.
Whereas for strongly coupled networks the value of the epidemic onset of the whole system is smaller than that of the
isolated networks (as in the SIS model), when the coupling is weak (i.e. the internetwork links are much less than those
connecting individuals within the same network), the system can sustain an epidemic state in one network while not
affecting the remaining ones. These mixed states point out that for the SIR dynamics there is a critical internetwork coupling,
below which the network behaves quite similarly to a set of unconnected networks.
5.2.3. Disease spreading on multiplex networks

As pointed out many times already, multiplex networks are, from the structural point of view, just particular cases of
multilayer networks in which each node is represented in each of the layers (so that the layers have the same number N
of nodes). In addition, each layer is connected to another one via N connections resulting from the links that each node
establishes with its representations in the remaining layers. Given that the results shown for the disease spreading on top of
an interconnected network are rooted in the spectral properties of its contact matrix, the case of a single disease spreading
on top of a multiplex network can be analyzed in the same way by considering that the (MN × MN) adjacency matrix is of
the particular form
   
A1 0 ··· ··· 0 0 I ··· ··· I
0 A2 0 ··· 0   I 0 I ··· I 
Â = A + C =  + . (216)
··· ··· ··· ··· · · · ··· ··· ··· ··· · · ·
0 ··· ··· ··· AM I ··· ··· I 0
However, a multiplex automatically sets the constraint that when a node gets infected by the disease in one layer, and
therefore it becomes a spreader of the disease in any of the other layers. This ingredient can only be approached (in the
interdependent framework) by setting β ab = β ba = 1 and β aa , β bb , µa , µb ≪ 1, which is a far different case with respect
to the weakly coupled regime usually studied in interdependent networks.
In fact, in Ref. [246] the Authors analyzed the SIS model in a multiplex in a similar way to Ref. [53], i.e. by writing the
microscopic dynamics. In this case, the expression for the probability that a node i is not infected by any neighbor (in network
A or B), qi (t ), reads
M 
N
(1 − β α (Aα )ij · pj (t )),

qi (t ) = (217)
α=1 j=1
where Aα is the adjacency matrix in layer α . Note that, at variance with Eq. (207), the vector containing the probability that
⃗, has N components since now the state of a node must be the same in any of the layers composing
any node is infected, p
the multiplex. One can note that the only difference with an epidemic spreading on the projected network (the network
resulting from the sum of the layers) is that the transmission rates β α are different for each of the layers. The analysis of the
microscopic dynamics relates the epidemic onset with the maximum eigenvalue of the matrix, defined as
M
(1 − β α (Aα )ij ).

Rij = 1 − (218)
α=1
Again, it is shown that the epidemic onset is always smaller than those of the isolated networks.
The SIR model in multiplex networks has been analyzed in Refs. [247,248]. In Ref. [247], the Authors made use of a two-
layers network in which a SIR disease switches layer at the expense of a cost. They derived the epidemic threshold, obtaining
a similar result to that of SIR dynamics in interconnected networks [245] when the coupling between networks is strong:
the epidemic threshold cannot be larger than those of the isolated networks. Obviously, multiplexity incorporates naturally
the constraint that the network layers are far from the weak coupling regime in which mixed states appeared. In this work,
the Authors derived an interesting result concerning the behavior of βc in a multiplex composed of two ER graphs as a
function of the difference, δλ, between the intralayer and interlayer transmission rates and the difference, δ z, between the
average degree of both layers. It was shown (see Fig. 36) that the multiplex epidemic threshold βc grows with δλ (when
the intralayer transmission rate increases with respect to the interlayer one), for low values of δ z. However, this trend is
reversed when δ z grows enough, so that βc decreases with δλ.
In Ref. [248], the SIR dynamics is addressed on top of a multiplex containing one physical layer and a second one in which
only a subset of the nodes in the first layer takes part. Importantly, the size of this subset is tunable so that they are able to
derive the conditions for the spreading of the disease as a function of the size of the second layer.
Competing epidemics. The particular setting of multiplex networks provides the best suited structural backbone for the
study of the dynamical interplay between different dynamical processes taking place within the same set of nodes. In this
way, multiplexity offers the possibility of incorporating different network layers for each dynamical process at work. This
possibility leads to approach a well studied problem in single-layer networks: the competition between the spreading of
two different diseases. In Ref. [249], Newman proposed a model in which two consecutive SIR diseases spread within a
single-layer network. For the first disease at work, the critical properties and the diffusion of the infections remain the
same, as usual. However, the second disease finds it more difficult to spread since those agents recovered from the first of
the diseases gain immunity also to the second one. The main result is that, even for SF networks, the epidemic threshold
for the coexistence of both diseases is larger than zero. Subsequently, other works extended the framework to the case of
simultaneous spreading of diseases [250], or to the SIS model [251].
The further extension of those works to multiplex networks is natural, and adds the possibility of exploring the effects
of having two different transmission networks for each one of the diseases, and thus incorporating the fact that different
diseases may have different transmission channels. Funk and Jansen were the first to address this issue in Ref. [126] by
considering, as in Ref. [249], that the disease spreads in a consecutive way. To this end, they considered a multiplex composed
of two networks and a SIR model so that the first spreading takes place in one of the networks leaving a set of recovered
nodes. These nodes are also set as recovered in the second layer (as if they were fully immunized) prior to the second
spreading that takes place in this second layer. The main result is that a positive correlation between the degrees of the
nodes in both layers enhances the immunization of the network to the second spreading (increasing the effective epidemic
threshold for the coexistence of two epidemics). This approach has been generalized in Ref. [127] to the scenario of two SIR
1 0.8
"25" u 2:1: 3
0.8 0.7
δz 0.6 0.6
0.4 0.5
0.2 0.4
0 0.3
0 0.2 0.4 0.6 0.8 1
δλ
Fig. 36. (Color online) The contour plot shows the epidemic threshold βc as a function of δ z and δλ for a two layer multiplex composed of two ER networks.
The dashed line (black) is a boundary of ∂βc /∂δλ > 0 for all δλ, and the green (blue) line indicates the lowest (highest) value of βc at a given δ z. βc shows
non-monotonic dependence on δλ.
Source: Reprinted figure from Ref. [247]. Courtesy of B. Min.
diseases whose spreading takes place simultaneously in the two layers of the multiplex, allowing to study the influence of
delay between two spreading processes and the effect of full and partial immunity.
The study of interacting epidemics has been also applied to the SIS model in Refs. [252–254]. In fact, the extension there to
multiplex networks leads to a modification of the standard SIS model into a SI1 I2 S model with the so-called mutual exclusivity,
so that when an agent is affected by one of the diseases, Iα , spreading in layer α , this automatically confers immunity to
the disease running in the other network layer. Obviously, once this agent becomes susceptible again, it may be infected
by any of the two viruses. The analysis can be tackled by writing the microscopic equations of the SI1 I2 S defined by two
adjacency matrices A1 and A2 and considering different disease parameters for each of the two spreading viruses: β1 (β2 )
for the rate of transmission of disease 1 (2) and µ1 (µ2 ) for the rate of the transition from I1 (I2 ) to the susceptible state. In
Ref. [252], the Authors made use of a coarse-grained microscopic approach shown in Ref. [255] for SIS, and that represents
an intermediate step between the accurate microscopic equations [235] and the HMF approximation. The set of equations
for the probabilities that agents are infected by the two viruses read as
N

p1i (t + 1) = p1i (t ) − µ1 p1i (t ) + β1 (1 − p1i (t ) − p2i (t )) (A1 )ij p1j (t ), (219)
j =1
N

p2i (t + 1) = p2i (t ) − µ2 p2i (t ) + β2 (1 − p1i (t ) − p2i (t )) (A2 )ij p2j (t ). (220)
j =1
As usual, this system of equations can be linearized around the trivial solution, pαi = 0 ∀ i, α , and the conditions under
which an epidemic in either population 1, population 2 or both will hold, can be derived. The conditions will depend on both
control parameters τ1 = β1 /µ1 and τ2 = β2 /µ2 . A trivial statement is that whenever τ1 < 1/Λ(A1 ) and τ2 < 1/Λ(A2 )
there is no epidemic in any of the network layers. However, a relevant scenario appears when any of the control parameters
(or both) is (are) above this threshold, since competition between the spread of the two viruses appears. In Ref. [252], the
Authors performed an analytical derivation based on perturbation theory and numerical simulations to derive the phase
diagram as a function of τ1 and τ2 . Analytical and numerical results are shown in Fig. 37, highlighting the existence of four
regimes: the free phase (N), epidemic in population 1 and free phase in 2 (I), epidemic in population 2 and free phase in 1
(II), and finally the coexistence of both epidemic states in layers 1 and 2 (III). Interestingly, from the diagrams, it is clear that
when the infection rate is large enough in one population, say 1, it hinders the epidemic onset in 2, even when τ2 > 1/Λ(A2 ).
A further relevant result of Ref. [252] is that whenever the two network layers are identical, the coexistence region
disappears.
Interaction between social dynamics and epidemic spreading. As mentioned above, the multiplex structure allows to couple
different dynamical processes involving the same set of agents being, in principle, different the structural backbone on
1
7 7
0.9
6 6 0.8
5 5 0.7
0.6
4 4
0.5
3 3 0.4
2 2 0.3
0.2
1 1 0.1
0 0 0
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
Fig. 37. (Color online) (Left panel) The survival regions diagram in the SI1 I2 S model in two-layer multiplex. Four possible scenarios appear: extinction
region (N) where both viruses die-out, mutual extinction region I, where virus 1 survives and virus 2 dies out, mutual extinction region II, where only virus
2 survives and virus 1 dies out, and finally the coexistence region III, where both viruses survive and persist in the population. The red arrow shows the
survival region of virus 1 (regions I and III) and the green arrow shows the survival region of virus 2 (regions II and III). (Middle and right panels) The two
contour plots show the results from numerical simulations of the SI1 I2 S model for a multiplex composed of two SF networks of N = 1000. The fraction of
infected individuals in the steady state of the dynamics in network 1 (middle panel) and network 2 (right panel) are shown. The white lines are theoretical
threshold curves accurately separating the survival regions as in the left panel.
Source: Reprinted figure from Ref. [252]. Courtesy of C. Scoglio and F.D. Sahneh.
top of which each process operates. In this way, disease spreading has been recently coupled to social dynamics [256],
in order to take into account how the social awareness about the spreading of a disease can influence the impact of the
latter. To this aim, the Authors coupled the SIS spreading taking place on a physical network with another compartmental
stochastic model in which the state of the agents can be of two kinds: Aware (A) and Unaware (U). The first dynamics is
completely characterized by the usual two parameters: the transmission, β , and recovery, µ, rates. The second dynamics
takes place on the second layer (a kind of virtual network in which information about the epidemic flows) and the transition
between states A and U is analogous to that in the SIS model: the direct contact of an A agent with another one in U
produces two A agents with probability λ, whereas an aware individual can become unaware with probability δ . The coupling
between the two dynamical models is done by reducing the transmission rates from infected individuals to aware and
susceptible ones by a factor γ (γ = 0 means that aware individuals become fully immunized). In addition, the state
unaware–infected is not admitted and, in a time discrete-version, it becomes aware–infected with probability 1 (if not
recovered).
As in Ref. [252], when isolated, the two dynamical processes display a critical point at some values of the respective
control parameters, β/µ and λ/δ . In particular: βc = µ/Λ(A1 ) and λc = δ/Λ(A2 ), where A1 and A2 are the adjacency
matrices of the first layer where SIS and Aware–Unaware dynamics take place respectively. Once again, the coexistence of
these processes changes the onset of the epidemic state. In particular, when the infection rate of the SIS dynamics is large
enough, the infected agents trigger the mechanism of awareness in the virtual layer, thus lowering the incidence of the
disease and increasing the epidemic onset. This result is shown in Fig. 38 for several values of µ and δ . Interestingly, as the
ratio δ/µ decreases, the value of the epidemic onset is more influenced by the Aware–Unaware dynamics.
Impact of interlayer topological properties. Except for the influence of dynamical processes located in other layers, the
network properties between different layers also play a significant role in the diffusion of disease. The first viewpoint comes
from the investigation of coupling density, which tries to reproduce the empirical fact that not all the nodes are present
on each layer of social networks. In Ref. [257], a fraction q of nodes are shared by two layers enabling to form partially
overlapped networks. Since there exists the same disease in both layers, the equal transmission probability is assumed,
whereas the spreading of infection starts from one random node. The main aim becomes here to explore the effect of the
overlapping fraction q on the disease propagation with reference to the SIR dynamics model. Based on the theoretical analysis
and simulations, the Authors found that the epidemic threshold of multiplex networks depends on both topology and the
coupling density. In the limit of q → 1, the epidemic threshold is minimum, and both layers have the identical ratio of
recovered nodes at any transmission probability. This is because nearly each node belongs to two layers at the same time.
However, with q decreasing, the epidemic threshold monotonously enhances, so as to reach the maximal value (at q → 0)
dominated by the isolated threshold of the layer with larger propagation ability. This point means that, even though there
exists negligible differences between the system composed of two isolated networks and the system connected by just a few
interlayer connections, these links can make the epidemic threshold of the isolated network with lower spreading ability
discontinuously change the threshold of the other layer. This finding is shedding light into the non-medical intervention for
the control of the disease propagation.
Of particular interest is distinguishing how the disease spreads if the nodes are coupled by a certain rule from the case
of simple overlapping. In a recent investigation [258], a two-route transmitted disease employing SIR model is introduced
into the multiplex framework, where two transmission routes take place on each layer, respectively. Moreover, the Authors
Fig. 38. (Color online) Dependence of the onset of the epidemics βc as a function of λ for different values of the recovery rates δ and µ for a two-layer
multiplex composed of: (i) a SF network of N = 1000 and exponent 2.5 in the physical layer, where SIS dynamics takes place, and (ii) the same SF network
of the physical layer plus 400 extra random links (non-overlapping with previous) in the virtual layer, where Aware–Unaware dynamics works. The shaded
rectangle shows the free phase (in which all individuals are Unaware and Healthy) when the layers are uncoupled. The boundaries are defined by the bound
of the structural characteristics of each layer: 1/Λ(A1 ) and 1/Λ(A2 ).
defined two new measures to evaluate the influence of interlayer links. The first is the degree–degree correlation (DDC)
between layers: positive (negative) values indicate that high (low) degree nodes are inclined to have more (less) connections
within another layer. The other quantity is the average similarity of neighbors (ASN) from different layers of nodes, which
measures how many joint neighbors one node possesses in both layers. The Authors found that, even if both layers are
below their respective thresholds, the disease can still propagate across the multiplex system. If the value of DDC is larger,
a lower disease threshold and a smaller outbreak size are supported. The epidemic threshold and disease size keep both
robust against the change of ASN.
5.2.4. Voluntary prevention measures

Except for the outstanding achievements of the prediction of the outbreak threshold, another task of utmost urgency,
during the study of epidemiology, is searching for the effective policy or strategy to control the epidemic spreading, and
to warrant public health [259,260]. This point seems particularly remarkable with worldwide pandemics of severe acute
respiratory syndrome (SARS), bird flu and new H1N1 flu [261–263]. A great number of immunization scenarios, such as
targeted immunization [264], ring immunization [265], and acquaintance immunization [266], have been proposed, but the
majority of them, based on the premise of compulsory requirements, completely neglects the individuals’ willingness and
desires, which instead follow many social factors including religious beliefs, the human rights and social total expenses
as well [267–270]. In this sense, voluntary prevention measures, incorporating both studies of infectious disease and
spontaneous human behavior traits, become popular programs [271,272], namely, the so-called ‘‘disease-behavior feedback
dynamics study’’.
Voluntary vaccination on single-layer networks. Among several existing control measures, vaccination attracts the hugest
attention from theoretical and experimental perspectives [273–276]. During the seasons of disease spreading, the high risk
of infection usually stimulates individual vaccination enthusiasm, while high expenses or side effects of a vaccine may lead
to the opposite case. Thus, individuals face a dilemma: to vaccinate or not, both of which ways can be managed as two basic
strategies with respective cost functions. To resolve this dilemma, game theory has been integrated into individual decision-
making processes, where nodes hold the prevention strategy according to risk and expenses in the epidemic campaign [277].
In a recent research [278], where nodes are only allowed to make rational selection (i.e. taking vaccination only if its expense
is low), the Authors found that disease outbreak can be more effectively inhibited on SF networks than on random networks.
Compared with early theoretical predictions of disease thresholds [239], this interesting outcome indicates that voluntary
vaccination can effectively control the spreading of diseases in real-world networks. Similarly, when the imitation dynamics
with noise replaces the deterministic rule in the disease-behavior study, multiple effects spontaneously arise: for perfect
vaccine, vaccination behavior of SF networks beats that of random networks, while imperfect vaccine enables the promotion
of vaccination and eradication of disease to act better on homogeneous connectivity patterns [279]. In Ref. [280] it is found
that the spatial population structure shows a ‘‘double-edged sword’’ effect, which fosters vaccine uptake at the low cost
of vaccination, but causes vaccination to plummet when the cost goes beyond a certain threshold. More specifically, if a
70
60
Attack/Adoption Rate (%)

50
40
30
20
Influenza–only Attack Rate
10 Coupled Diffusion Attack Rate
Coupled–Diffusion Adoption Rate
0
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Transmission probability
Fig. 39. Attack or adoption rate vs. the transmission probability in disease-only diffusion processes and coupled diffusion processes, where influenza is
selected as the potential disease. Here, the adoption rate denotes the percentage of nodes adopting prevention behavior.
© 2012, with permission from Elsevier.
small fraction of individuals consistently holds vaccinating in flu-like diseases, these committed vaccinators can efficiently
avoid the clustering of susceptible individuals and stimulate other imitators to take vaccination, thus contributing to the
promotion of vaccine uptake [281].
Voluntary prevention behavior on multilayer networks. Although both infection diseases and prevention measures diffuse
simultaneously through networks and interact with each other, few existing studies have coupled them together. Recently,
the research of disease-behavior interactions on top of multilayer structures has also received attention. In the first
conceptual work [282], two diffusion processes are implemented in the multiplex viewpoint, where two networks, which
are fully or partially coupled via the interlayer links, involve the same individuals yet different intralayer connections. Under
such a framework, one network is regarded as the infection network to transmit diseases, and the remaining communication
network channels individual influence concerning preventive behavior. Moreover, at variance from the aforementioned case
where vaccination is the principal measure on single networks, other prevention behaviors, such as wearing facemasks,
washing hands frequently, taking pharmaceutical drugs, and avoiding contact with sick people, are considered as the main
strategies, which seem particularly reasonable before attaining effective vaccine [283].
Based on the online surveys, the threshold model is selected, whereas only once the proportion of adopters among
intralayer neighbors of communication networks reaches the perceived threshold of adoption pressure (or the proportion
of infected individuals exceeds the perceived risk of infection on the infection network), the node can adopt the preventive
behavior. Surprisingly, the Authors found that, as compared to the widely used disease-only model, the coupled diffusion
model leads to a lower frequency of infection. Fig. 39 shows that, as preventive behavior protects a certain people from being
infected, the threshold and peak of disease outbreak are effectively suppressed in the new embedded scenario. Furthermore,
the sensitivity analysis identifies that the structure of infection network has dramatic influence on both epidemic and
adoption percentages, while the variant of communication network produces less effects.
After such a seminal research, the role of multiplex networks in the coupled disease-behavior system has been further
investigated. Recently, a triple diffusion process, comprising disease transmission, information flow regarding disease and
preventive behaviors against disease, was integrated into multilayer social networks [284]. Among the interaction of the
three diffusion dynamics, the negative and positive feedback loops are formed. Here it is worth mentioning that, as compared
with Ref.[256] involving the information diffusion as well, the effect of prevention behavior is of particular evidence. Under
the help of agent-based method, this new conceptual framework can reasonably replicate the observed trends of influenza
infection and online query frequency in the empirical population.
Finally, a similar scenario was proposed in Ref. [285], where the social and biological dynamics processes are embedded
into a multiplex viewpoint. Under such a disease-behavior system, biological contagion spreads on one layer, while the
main control measure, vaccination, is carried out in the social network. At variance with the case of an isolated network, the
Authors suggested that the resulting coupling can bring positive or negative impact on the vaccination coverage during
a vaccine scare, i.e. accelerating or decelerating the elimination of disease, which can get support from the empirical
examples. Due to the variety of factors on the social layer, the epidemic trajectories and the uptake of control measures
can also change to a large extent. Moreover, inspired by these findings, the Authors suggested that the coupled disease-
behavior model, which gets empirical validation, can be used as a predictive tool to explore the optimal strategies for public
health.
5.2.5. Opinion formation dynamics

Understanding the emergence of a macroscopically ordered state is one of the based questions in natural and social
sciences. To elucidate this issue, dozens of experimental and theoretical frameworks have been proposed from viewpoints
of statistical physics, economics and sociology [227,286–289]. Opinion formation dynamics, as one simple and paradigmatic
dynamical process, has attracted particular attention. Up to now, various spin-like setups have been put forward to study the
evolution of opinion dynamics, which include the Sznajd model [290], majority rule model [291,292], bounded confidence
model [288,293], and voter model [294,295], to name but a few. Among the existing scenarios, the voter model is attracting
the most remarkable attention [286,296], and therefore it is also the main candidate in this subsection.
The basic definition of the voter model is rather simple: each agent (voter) is endowed with a binary variable s = ±1.
At each time step, an agent i is chosen along with one of its neighbors j and si = sj , i.e., the agent adopts the opinion of
the neighbor. In this sense, the transition rate of individual opinion is proportional to the fraction of neighboring nodes
in the opposite state. The dynamical process is repeated until the states with all sites equal (consensus) are absorbing. In
spite of being extremely simple, the voter model produces rich dynamical behaviors [295,297], especially from the angles of
spatial interaction. In one-dimensional lattices, the dynamics is exactly the same as the zero-temperature Glauber dynamics
[298]. Turning to two-dimensional lattices, the growth of domains can makes the interfaces quite rough, at odds with usual
coarsening systems [299]. Moreover, a series of extended versions, such as the vacillating voter model [300], nonlinear
voter model [301], heterogeneous voter model [302], and constrained voter model [303] have also been suggested. In the
following we will present more details on the voter model.
Voter model on single-layer networks. As it is well-known, the property of spatial topology usually has a nontrivial influence
on the dynamical models. In this sense, the ordering dynamics of voter model should not be exceptional. If employing
one complete graph, the mean consensus time ⟨T ⟩ can be easily obtained by using the Fokker–Planck equation [304]. The
consideration of a heterogeneous connectivity can make the mean consensus time ⟨T ⟩ to become closely related with the
structural network details. On SF networks with exponent γ , direct voter model makes ⟨T ⟩ to scale as the size N of the
network for γ > 3 and sublinearly for γ ≤ 3 [296,305,306]. Non trivial relationships between ⟨T ⟩ and N are also found in
other versions of the voter model as in Ref. [307]. For small-world networks, the plateau formed by active interfaces hinders
the diffusion of opinions, so that the consensus eventually relies on the linear system size [308,309]. Along this research
line, Zschaler et al. recently examined the effect of directed networks on the order–disorder transition [310]. The impact of
other spatial mechanisms, such as social influence and heterogeneous beliefs, are also of particular evidence [311,312].
At variance with the aforementioned processes on static networks or in a Markovian memoryless fashion, coevolution
voter model has also been duly studied. We first focus on non-Markovian patterns. In recent researches [314,315], Stark et al.
explored the effect of a memory-dependent transition rate, which is determined by the persistence time of individual current
opinion. The longer a voter keeps its opinion, the less probable is that it switches the status in the long run. The Authors found
that the process toward reaching a macroscopic consensus was accelerated by slowing down the microscopic dynamics.
Looking at another recent report [313], where individual ability q of opinion switching is weakened during a freezing period
H, the Authors revealed that an intermediate freezing period can lead to the fastest consensus, i.e. to the smallest ⟨T ⟩. The
essence of this accelerated consensus is attributed to the biased random walk of the interface between adjacent opinion
clusters (as shown in Fig. 40). Moreover, it is worth checking how the dynamical topology affects the diffusion of opinions.
Under such a framework, an agent can adjust its connections with a probability (otherwise it updates its opinion). The state
of the system on coevolving networks can be characterized by the interface density ρ quantifying the fraction of edges
linking nodes with different states (active links). When ρ = 0 the system is in the active phase, while for ρ ̸= 0 it keeps the
frozen phase. Based on the mean-field approximation approach, Ref. [316] revealed that the competition between opinion
imitation and the rewiring links drives a fragmentation transition from the active to the frozen phase.
To mimic some social phenomena and check voting phenomenology, many modifications have also been suggested
based on the original voter model. The most renowned extension is the presence of the ‘‘zealot’’, who never changes
its opinion [317]. For low-dimension networks, this type of agent can influence the selection of all the agents, and lead
to the general consensus in the form of its own opinion. However, once higher-dimension networks are considered it
becomes rather difficult to inspire the consensus in the whole system, but the ‘‘zealot’’ just induces a partial bias in the
local neighborhood [302,318,319]. In addition, the voter model has also been applied to biological and ecological problems,
where people try to protect the diversity of species and fixation in the evolution of competing species [320,321].
Voter model on multilayer networks. With the fast development of network science, the research of opinion dynamics has
also been extended to multilayer networks. One potential approach is to create interlayer links, through which opinions
can be exchanged between different networks. In a recent research [322], the Author considered the case of a two-clique
graph, which is composed of 2 cliques interconnected by m interlayer links. If the size of both cliques is N, each node
has m/N interlayer links on average. These links are either regularly placed, or distributed at full random. At each time
step, one link is randomly selected with equal probability 1/[N (N − 1) + m], and then the opinion imitation takes place
between two endpoints of this link. In this sense, the impact of the number of interlayer links per node becomes of
particular interest. Through strictly mathematical analysis and computational simulations, the Authors unveiled the scaling
relationship between the time to consensus and the number of interlayer links per node. If there exist many interlayer links
per node, the mean consensus time is ⟨T ⟩ = O(N ), i.e. the system reaches the fast consensus regime, which agrees with the
a (×10 4) H=0 b (×10 3) H=15 c (×10 4) H=4000
0.2
0.5
0.5
0.4
1.0
1.0
0.6
1.5
0.8
1.5
t 2.0
1.0
2.5 2.0
1.2
3.0
1.4 2.5
3.5 1.6
3.0
Fig. 40. (Color online) Snapshots of the evolution of opinion dynamics with three typical freezing periods H. From left to right, the values of H are 0, 15
and 4000. In each panel, the vertical axis corresponds to the elapsed time, while the yellow and blue areas on the horizontal lines represent the opinion
clusters. The total time spent to reach a global consensus, i.e. the consensus time ⟨T ⟩, under the scenario of intermediate value (H = 15) is far less than
two other cases. The size of networks is N = 500, q = 0.01 is the switching ratio during the freezing period. (For interpretation of the references to colour
in this figure legend, the reader is referred to the web version of this article.)
Fig. 41. (Color online) Relationship between the normalized mean consensus time, ⟨T ⟩/N, and the number of interlayer links per node, m/N, for different
network sizes. It is clear that there exist two consensus regimes. The solid line represents the relationship ⟨T ⟩/N ∝ (m/N )−1 as guides to the eye.
Source: Reprinted figure from Ref. [322]. Courtesy of N. Masuda.
expectation of a complete graph [307]. However, once the number of interlayer links per node is much smaller than unity,
the mean consensus time is ⟨T ⟩ = O(N 2 ), i.e. the system attains the slow consensus regime. The crossover between these
two regimes takes places at approximately one interlayer link per node, as shown in Fig. 41. Thus, these results suggest
that the sparse interlayer connections between two networks extremely decelerate the ordering state for the voter model
dynamics.
In Ref. [323], a coevolution voter model is incorporated into a multilayer framework, where a fraction q of nodes has
the interlayer links across two networks. On each layer l (l = 1, 2), individual opinion and intralayer link are allowed to
evolve in accordance with their own topological temporal scales. Node i selects one random neighbor j in the same layer.
If both have different opinions, node i copies the state of j with probability pl , otherwise (with probability 1 − pl ) it severs
the connection with j and draws a new link to one randomly chosen node in the same layer possessing a state identical
to itself. The case p1 = p2 preserves the symmetric multilayer setup, namely, the time scales are the same in both layers.
At variance, for p1 ̸= p2 , an asymmetric multilayer case takes place, namely, different topological time scales characterize
the two layers. In the symmetric case, the Authors found that interlayer connectivity plays a critical role in determining
the state transition of systems. Once again, the absorbing fragmentation transition appears at a certain threshold value of
pl . Moving to the asymmetric case, the Authors unveiled an anomalous shattered fragmentation transition, where the layer
dynamics splits into many isolated components possessing different opinions, namely an explosive growth of the number
of disconnected components. Moreover, Ref. [323] also identified the critical degree of interlayer connectivity to stop the
fragmentation of one layer by coupling it to another layer that has no fragments.
Finally, it is interesting to consider the application of opinion dynamics in the multiplex perspective. In a recent
research [324] a simple model of opinion formation dynamics is proposed to describe the parties competing for votes within
a political election. At variance with previous setups of voter model [286,296], each opinion or party is modeled here as
one social network where the contagion dynamics takes place. Every agent, at the election day, can choose either to be
active in only one of the two networks (vote for one party) or to be inactive, namely, not to vote on both networks. Based
on the simulated annealing algorithm, the Authors showed a rich phase diagram. In a wide region of the phase diagram,
the competition between the two parties allows for the emergence of pluralism, namely two antagonistic parties gather a
finite share of the total votes. The densely connected network usually wins the election campaign, regardless of the initial
distribution. Nevertheless, small perturbations could break the existing equilibrium, and a small minority of committed
agents can alter the election outcome, which is consistent with recent claims [325].
5.3. Evolutionary games on multilayer networks
The study of evolutionary games on top of single-layer networks has attracted a lot of interest as a connector between
statistical physics, evolutionary dynamics and social sciences [326–328]. This is a very appealing research topic since the
dynamics implemented in the network arises from an evolutionary approach [329]. In addition, the study of evolutionary
games on networks has been one of the most studied avenues to understand the emergence of cooperation in different
contexts [330]. The emergence of such a collective phenomenon, cooperation, is a key issue that arises when studying
seemingly diverse evolutionary puzzles, such as the origin of multicellular organisms [331], the altruistic behavior of humans
and primates [332], or the way advanced animal societies, such as ant colonies, work [333,334], among others.
Previous research on evolutionary games on networks has mainly focused on social dilemmas, in which agents can
take one of two strategies: Cooperation (C) and Defection (D). Among these dilemmas, the Prisoner’s Dilemma game (PDG)
represents the most addressed paradigm when studying the emergence of cooperation [335,336]. This game is fully
described by the following payoff matrix
 C D   C D 
C R S C 1 0
5= = , (221)
D T P D b>1 1>ϵ≥0
showing that when two cooperators meet, each one obtains a ‘‘Reward’’ (R = 1) whereas the interaction between two
defectors causes a ‘‘Punishment’’ (P = ϵ ) to both. However, cooperation is hampered by the players’ ‘‘Temptation’’ to defect,
since defecting against a cooperator yields more payoff than cooperating (T = b > 1), and by the risk arising from
cooperating with a defector, as it yields the lowest payoff, usually called ‘‘Suckers’’, (S = 0) [337]. These payoffs satisfy
the ranking T > R > P ≥ S. The social dilemma is thus established, since mutual cooperation yields both an individual and
total benefit higher than that of mutual defection.
The pairwise nature of the game is translated to a population scale by making the agents interacting (playing) with
each other, and accumulating the payoff obtained from each interaction. After each round of the game, the strategies of
the agents are updated so that those agents with less payoff are tempted to copy (replicate) the strategy of those fittest
individuals. In unstructured populations, in which players are well-mixed, evolutionary dynamics leads all the individuals
to defection [338,339]. However, the existence of a network of interactions, so that each agent can only play with those
directly connected to it, the population can sometimes promote the emergence of cooperation. This mechanism promoting
cooperation was coined as network reciprocity [340], and it was observed to be substantially enhanced when the network
substrate is SF [341,342].
Very recently [328] the attention of the community has been extended to the n-player (n > 2) generalization of the PDG,
also called Public Goods Game (PGG) [343]. Under this setting, each cooperator contributes a fixed investment v to the public
good whereas defectors do not contribute. The total contribution of the n agents is then multiplied by an enhancement factor
1 < r < n and finally the result is equally distributed between all n members of the group. The payoff of cooperators and
defectors in a group with a fraction x of cooperators is, respectively
fC (x) = (rx − 1)v, (222)

fD (x) = rxv. (223)
Thus, defectors obtain the same benefit of cooperators at zero cost, i.e. they free-ride on the effort of cooperators. As in
the PDG, the evolutionary outcome of the PGG differs if played on a well-mixed population, in which the dynamics ends
up in a full defector population, or on a structured one. In particular, Brandt et al. [344,345] showed that, when games are
restricted to local interactions, cooperation is promoted so that full cooperation is observed for values of the enhancement
factor r < n. This result was later generalized to complex networks [346] and bipartite graphs [347,348].
5.3.1. Pairwise interaction games
Games on interconnected networks. The study about emergence of cooperation in the PDG in networked populations has
been also extended to the case of multilayer systems. One of the approaches is to consider that each layer is composed of a
different set of agents, so that each of them plays both with those neighbors within the same network layer and with those
placed in the other ones [349,350]. In Ref. [349] the Authors considered the case of 2 interconnected networks in which
each agent plays a different game depending on whether the neighbor belongs to its same network, or to the other one.
In particular, they considered two possible payoff matrices, one for those interactions between nodes of the same network
layer, 5intra , that takes the same form of Eq. (221), and the following one for interlayer interactions
 C D 
C 1 0
5inter = . (224)
D b>1 ϵ<0
Thus, if A1 and A2 are the adjacency matrices of the two networks and C12 = C21 is the matrix containing the interlayer
connections (so that the total adjacency matrix reads as in Eq. (205)), the payoff of a node i belonging to network layer
α is
N1 N2
fiα (t ) = (Aα )ij · s⃗i α (t )T 5intra s⃗j α (t ) + (C12 )ij · s⃗i α (t )T 5inter s⃗j β (t ), with α ̸= β,
 
(225)
j=1 j =1
α α
where s⃗i is a two-components vector denoting the strategy of the node i in layer α at time t: s⃗i reads (1, 0)T for Cooperation
and (0, 1)T for Defection.
Thus, nodes gain payoff according to a PDG when playing with neighbors in the same layer, whereas for interlayer
connections the payoff remains the same except for those interactions between two defectors, as the Punishment becomes
negative being the smallest possible payoff in this matrix, i.e. the payoff ranking evolves into T > R > S > P. This
latter change turns the PDG into a Snowdrift game (SDG). In this way, the exploitation that defectors make of cooperators,
regardless of the layer where they are located, is hampered by the payoff loss when meeting a defector belonging to the
other network layer. The results in Ref. [349] reveal that, both for well-mixed layers and networked ones, it is possible to
find polarized states in which cooperation dominates in one of the layers whereas defection does in the other. Fig. 42 shows
the fraction of cooperators, c, as a function of the Temptation b. When there are no connections between network layers
(p = 0), so that the two populations play a PDG, the diagram follows the usual trend in single-layer networks: cooperation
resists for small values of b > 1 and then suddenly decreases so that c = 0 for b ≫ 1. However, polarized states (here
revealed by the plateau at c = 0.5) appear as soon as the two layers are connected, i.e. even for a small number of interlayer
links, thus showing the importance that the interconnection between layers has on the evolutionary outcome.
In Ref. [350] the Authors addressed a similar interconnected system of 2 networks, but considering that both intra and
interlayer interactions are governed also by the same Eq. (221), i.e. the payoff matrix of the PDG. In this case, the Authors
focus on the emergence of cooperation as a function of the density of interlayer links and the topology of each layer at
work. In general, for low density of interlayer connections, the overall cooperation of the system is enhanced. In this regime,
it was found that those agents connecting the two network layers are more likely to cooperate than those sharing only
intralayer links. Furthermore, although in general the interconnection of layers promotes cooperation in the whole system,
when looking at the level of single layers, the effect of interconnection can produce a reduction of cooperation in one of the
layers that is compensated by the increase in the other one.
Aside from involving payoff via playing games with nodes of other networks, another potential approach is to allow direct
information transmission or exchange between layers. In a recent research [351], where PDG and SDG are respectively
implemented on different layers, a biased imitation is proposed: a node with probability p can mimic the strategy of
intralayer neighbor in the same network, otherwise the internetwork neighbor is chosen as the strategy donor. Under such
a framework, the Authors try to explore how the cooperation trait varies as a function of a biased probability p. The biased
probability unveils a multiple effect, which is robust against the change of population structures. A slight decline of the
biased probability starting from p = 1 promotes the cooperation behavior in the PDG layer. On the contrary, this decrease
impairs the evolution of cooperation in the SDG subpopulation. Ref. [351] also verifies that qualitatively valid observations
can be extended to a much broader parameter space in virtue of the mean-field method.
Fig. 42. (Color online) Fraction of cooperators c in the population 1 as a function of b, for different values of the fraction p of interpopulation contacts.
The population 1 (of size N1 = 103 ) has been coupled to a smaller population 2 (N2 = 102 ). While initial strategies in population 1 are equally probable
(random initial conditions), the population 2 starts from the absorbent state of fully defection. Other parameters are r = 0, ϵ = −0.4. Both populations
have a random (Erdős–Rényi) network of contacts with average degree ⟨k⟩ = 6.
Games on multiplex networks. A second stream of works addressed the study of evolutionary dynamics in multilayer
networks [352]. In this way, all the M layers have the same number of nodes, say N, since each node is the representation
of a particular agent, and the only links between layers are those connecting the nodes representing the same agent.
The assumption of a multiplex imposes dynamical correlations between the states (or strategies) of the replica nodes. In
Ref. [352] this correlation is implemented by considering that an agent i can, in principle, take different strategies in each
of the layers. For instance, in the PDG an agent i can be a defector in n layers, while it cooperates in the remaining M − n
layers. The total payoff of a node i in layer α at time t reads
N
α α
fiα (t ) = Aαij s⃗i (t )5s⃗j (t ),

(226)
j =1
where the payoff matrix is defined as in Eq. (221). In addition, the total payoff of agent i is the sum of all the payoffs
accumulated by the different representations across the M layers
M
fiα (t ).

fi (t ) = (227)
α=1
In Ref. [352] the update of the strategy used by an agent i at layer α depends on fi (t ) (instead of taking into account only
the payoff accumulated at this layer fiα (t )), thus coupling the evolution of the states of the nodes in a given layer with the
dynamical states of the rest of the layers. Fig. 43 shows that the correlation between layers influences the cooperation levels
of the system in a way that, by increasing the number M of layers, the resilience of cooperation is enhanced at the expense
of decreasing the level of cooperation observed for small values of the temptation b to defect.
Moreover, motivated by the fact that each agent is usually member of different networks in human societies, a completely
distinct setup is suggested in Ref. [353]. The system is there composed of 2-layer SF networks. One network, named
interaction layer, is used for the accumulation of payoffs, and the other serves as the updating network for strategies or states.
The Authors show that, with all the possible combinations of degree mixing, breaking the symmetry through assortative
mixing in one layer and/or disassortative mixing in the other as well as preserving the symmetry by means of assortative
mixing in both layers, impedes the evolution of cooperation. The essence of this scenario for the cooperation inhibition can be
attributed to the collapse of giant cooperation clusters, especially as compared to the case of single-layer SF networks [353].
Games on interdependent networks. The third realm of the evolutionary game in multilayer networks turns to that of
interdependent networks. It is worth mentioning that this form possesses evident differences with the aforegoing situations.
As compared to the case of interconnected networks, the direct interaction of game and strategy exchanged between
network layers are strictly forbidden. On the other hand, the corresponding nodes connected via the internetwork links
represent different agents, at variance with the multiplex scenario, thus both the accumulation of payoff and strategy
updating are just limited to the same network layer. In this sense, the internetwork links are mainly used to create mutual
Fig. 43. (Color online) Cooperation diagrams of multinetworks. Average level of cooperation ⟨c ⟩ as a function of the temptation b to defect for several
multinetworks with different number of layers M (the number of layers in indicated in the legend). In panel A the network layers are ER graphs with
⟨k⟩ = 3 (sparse graphs) while in panel B we have ⟨k⟩ = 20. In both cases N = 250 nodes. As it can be observed, the resilience of cooperation increases
remarkably as the number of layers M grows.
influence between different networks. Since utility and strategy are prime elements in evolutionary game theory, creating
interdependence through both factors becomes a natural choice. The framework of interdependent network games is then
the following: agents of each layer do not engage in any game interaction beyond their own layer, but just refer utility or
strategy information of other systems during the process of strategy updating with their intranetwork neighbors.
A recent research suggests the case of 2 interdependent networks, network α and network β , in which each node obtains
the payoff by playing the games with its intralayer neighbors [354]. Then the interdependent correlation between networks
(the node-to-point interdependence) is realized by defining an utility function as follows:
β
Uiα (t ) = (1 − γ ) · fiα (t ) + γ · fi′ (t ), (228)
β
where fiα (t ), fi′ (t ) represent the payoffs of a given node i in network α and that of corresponding partner i′ from network β ,
and γ ∈ [0, 1] is the parameter accounting for the coupling strength between layers. The function Uiα (t ) is then used for the
update of the strategy of agent i at layer α . Such an expression is a symmetric function, i.e. the payoff of the node in network
β also accounts for the same part (1 − γ ) in its utility. It is worth noting that setting γ = 0 decouples both networks, and
the model gets back to the traditional single-layer network case [326]. With the increment of the coupling correlation, the
utility of one network depends more and more on the other system. In particular, for γ = 1, the utility of one node becomes
fully determined by the other network, and the updating strategy becomes the process of tossing a coin.
The Authors selected the PDG as the paradigmatic model to explore the effect of the coupling strength. They found that
there exists a threshold value γt (γt ≃ 0.5) in the system, such that, when the coupling strength γ is smaller than this
threshold, the symmetry of cooperation level in both networks is maintained. However, as the value of γ exceeds γt , there
is a spontaneous symmetry breaking of the fraction of cooperators between different networks. Fig. 44 shows the typical
cross sections of phase diagrams for different coupling degree γ . In the case of γ = 0.4, where cooperation level is well
promoted, one can observe the existence of three sections: pure cooperators section, mixed strategies section and pure
defectors section. If γ is set as 0.9, a novel section emerges: the symmetry breaking section. Moreover, this observation is
universally effective on different networks and can be qualitatively predicted by the strategy-couple pair approximation
approach, in which an accurate prediction of the phase transition is difficult, due to the fact that long-range interactions are
neglected among nodes in the network [326].
5.3.2. The public goods game

The first work on this game in multilayered networks [355] addressed a simple system composed of two regular lattices
that are coupled in the usual coupling way (each node in a layer has a link with its corresponding partner in the other one).
The PGG is played according to the usual rules, so that in each of the layers, an individual constitutes a group together with
its 4 nearest intralayer neighbors, thus taking part simultaneously in 5 different PGGs. As described for PDG, the coupling
between layers comes from an utility function that couples the payoff obtained by the agent i in layer α , fiα , with that
Fig. 44. (Color online) Cross sections of phase diagrams for γ = 0.4 (left) and γ = 0.9 (middle) on the interdependent square lattices (green squares) and
small-world networks (red triangles) with a fraction of rewired links equal to 0.05. The colored regions correspond to different sections: yellow is pure
cooperators section, cyan is mixed strategies section, gray is pure defectors section and purple is symmetry breaking section. The lines in right panel denote
the results of strategy-couple pair approximation for γ = 0.9, which is qualitatively similar to the case of the middle panel. The dashed line represents an
unstable solution of the analytical approach. Note that different ranges for the horizontal axis are used in the three panels, just for the guide of the eye.
β
obtained by its partner i′ in the other layer β , fi′ . In this case, the coupling takes the form
β
Ui = γ · fiα + (1 − γ ) · fi′ . (229)
Here the utility function, at variance with Ref. [354], is the same for the two layers, and it incorporates a new ingredient since
there is one dominating layer depending on the value of γ (namely, a biased master–slave interdependence fashion): when
γ < 0.5 (γ > 0.5) layer α (β ) is dominated by layer β (α ). As usual, the utility function is used for updating the strategies
of the nodes in each layer. The most interesting result obtained in Ref. [355] is that, when one layer almost dominates the
other, cooperation is extraordinarily favored in the slave layer. For γ → 0 and γ → 1, the corresponding master layer
behaves almost equally to an isolated square lattice showing more weakness to defection than the slave one, in which the
relation between the strategy and the utility function is almost random. These results are explained in terms of time-scale
separation for the spread of strategies that favors the development of cooperator clusters in the slave layer.
Many other kinds of interlayer couplings can be introduced in the utility function. In Ref. [356], using a two-layer topology
composed of two identical 2-dimensional lattices, the Authors followed a similar approach to that used in Ref. [354] for the
PDG, since each one of the layers has its own utility function, which in its turn depends on what is the state of the other
layer. In this case, the utility function for the agent i in layer α is defined as
   
N N
α α α α β β
 
Ui = (1 − γ − γ ) · fi + γ ·
′
Aij fj +γ ·
′
Ai′ j′ fj′ , (230)
j =1 j ′ =1
where the terms multiplied by γ and γ ′ account for the average payoff of the neighbors of i in the same layer α and that
of those neighbors of the corresponding partner i′ in the other layer β , respectively. The effect of the coupling parameters
γ and γ ′ is the promotion of cooperation although this effect is more pronounced for γ than for the interlayer coupling
γ ′ (see the left panel of Fig. 45). This observation was given the name of interdependent network reciprocity. The essence
of this mechanism requires the simultaneous formation of correlated cooperator clusters on both networks, which can
get further support from the synchronized evolution of strategy pairs between networks, as shown in the right panel of
Fig. 45. If synchronization does not emerge or if the coordination process is disturbed, interdependent network reciprocity
fails, resulting in the total collapse of cooperation. Moreover, this promoting effect is proven to take place also in systems
involving more than 2 layers, as long as the coordination between the networks is not disturbed.
Another way to model the interlayer coupling via the utility function is the probabilistic interconnection introduced in
Ref. [357], in which two lattices are coupled in an interconnected way. However, two nodes of both layers are effectively
coupled with probability p, so that not all the possible N pairs of interlayer links are active. When an interlayer link is
active, the utility function of player i in layer α is influenced, so that apart from the 5 PGG in which it is enrolled within
the corresponding layer α , it also participates in that centered on its partner i′ in the other layer β . In this way, the Authors
showed that there is an optimal interlayer connection probability p that maximizes the cooperation level.
The information of strategy also plays a crucial role in affecting the trait of other systems. In Ref. [358], the Authors
proposed an alternative coupling between two lattices. The coupling is done via the update rule by which the agents decide
their strategies (to cooperate or to defect) for the next round of the game instead that via the utility function. The probability
that agent i in layer α adopts the strategy of a neighboring node j in the same layer depends here on the difference between
their payoff functions and on the fact that the corresponding partner of i in layer β has the same or a different strategy. If the
strategies of both the node i and its partner i′ are the same, the probability of changing is decreased with respect to the usual
one when the network layers are isolated. This coupling is seen not only to enhance cooperation in both PDG and PGG, but
also to sustain the spontaneous emergence of correlated behaviors between the two networks. If the range of information
sharing is extended to groups, this promotion effect is even more evident.
Fig. 45. (Color online) The contour plots show the stationary fraction of cooperators ⟨c ⟩ as a function of the coupling parameters γ and γ ′ entering in Eq.
(230) for r = 3.5 in PGG (left panel). Time courses of strategy pairs within and between networks (right panel), where Pi and Pe denote the corresponding
pair configuration probabilities within (internal) and between (external) networks, respectively. The parameters are r = 3.5, γ = 0.15 and γ ′ = 0.5 in
right panel.
5.3.3. Partial overlap and spatial coevolution

While the effect of interlayer nodes has been extensively identified, the question of what would be the best overlapping
level for the resolution of social dilemmas remains unclear. To answer this issue, one scenario of partial overlapping is
suggested in Ref. [359], where two networks possessing the same number N of nodes compose the interdependent system.
Before any interaction, a fraction ρ of the nodes of one layer, ρ N nodes, which are chosen uniformly at random, are linked to
their corresponding partners in the other one, and the residual nodes have no internetwork partners. Due to the existence
of interdependence, the utilities of these overlapping nodes are not simply payoffs obtained from the interactions with the
intralayer neighbors in the same network, but they rather involve the performance of their partners in the other network.
Here the correlation between the payoffs obtained by a given node i and its partner i′ is done by defining the following utility
function:
β
Uiα (t ) = fiα (t ) + γ · fi′ (t ), (231)
with γ ∈ [0, 1] determining the strength of coupling between layers. In particular, if the node i has no external links, its
utility simply equals to its own payoff (Uiα (t ) = fiα (t )). The utility function Uiα (t ) is then used for the update of the strategy
of agent i at layer α . As in the setup of Ref. [354], the utility function has here a symmetric form.
Under such a scenario, cooperation is studied as a function of ρ and γ . The Authors found that an intermediate density
of sufficiently strong coupling between networks (more accurately, γ → 1 and ρ ≃ 0.5) warrants an optimal regime
for cooperation to survive. This observation is robust to the variations of the strategy updating dynamics, the interaction
topology, and the governing social dilemma, thus suggesting a high degree of universality. Similarly to Ref. [350] it is
observed that, within a layer, cooperation tends to fixate preferentially in those nodes that are connected with their
counterpart in the other layer (see Fig. 46).
Furthermore, it is instructive to examine how the cooperative behavior survives on the coevolution framework, in which
the heterogeneity is usually ubiquitous [360,361]. In a recent research [362], where no interlayer links exist between two
independent structured populations, a utility threshold is designed to manage the emergence of temporary interlayer links.
If the current utility of node i exceeds this threshold, it is rewarded by one interlayer link to the corresponding partner
belonging to the other network in the next time step; otherwise the external link is not formed. As for the evaluation of
utility, Eq. (231) is retained. It is easy to estimate that, as long as the utility of all nodes drops below the threshold, the
external link is terminated, which returns to decoupled single-layer networks. The Authors showed that an intermediate
threshold gives rise to distinguished players that act as strong catalysts of cooperative behavior in both PDG and PGG.
However, there also exist critical utility thresholds beyond which distinguished players are no longer able to percolate. The
interdependence between the two populations then vanishes, and the survival of cooperation becomes a hard task. Finally,
this demonstration of the spontaneous emergence of optimal interdependence by means of a simple coevolutionary rule
based on the self-organization of reward and fitness is proven to be universally effective on different interaction networks.
At variance with the scenario of utility threshold, another setup referring to coevolution between strategy and network
structures considers the strategy transmission ability [364]. Initially, each node i of two independent networks has a low
and identical transmission ability wi . If node i succeeds in passing its strategy to one of its intranetwork neighbors, its
ability enhances by ∆, wi = wi + ∆; otherwise wi becomes wi − ∆. A threshold wth is selected as the criterion of creating
an external link: only when the transmission ability satisfies wi ≥ wth , player i is allowed to have an interlayer link to its
Fig. 46. (Color online) The contour plots show the stationary fraction of cooperators ⟨c ⟩ as a function of the coupling ratio ρ and coupling degree γ for
different cases. Both top panels and the left bottom one refer to a prisoner dilemma underlying game, whereas the snowdrift game is used in the bottom
right panel. Square lattices are used except for the bottom left panel, which uses triangle lattices. Similarly, the top right panel uses proportion imitation
rule [341,342], while Fermi update rule [326,363] is held in the other ones. It is obvious that the proposed scenario of partial overlap is universally effective.
corresponding partner in the other network. If wi drops below the threshold wth , the external link is terminated. The making
and breaking of the interlayer links between the two networks thus serves as individual rewarding and punishment, that
reflect the evolutionary success within a network. Here, the utility of a node with (without) interlayer link is in accordance
with Ref. [359], yet using the fixed strength of coupling. In this sense, the question becomes what are the optimal conditions
for resolving social dilemma.
The Authors showed that, due to coevolutionary rule, the interdependence between networks self-organizes so as to yield
optimal conditions for the evolution of cooperation even under extremely adverse conditions. In this case, approximately
half of the nodes form an interlayer link, while the other half is denied the participation in the activities beyond their host
network. Moreover, due to the spontaneous emergence of a two-class society, with only the upper class being allowed to
control and take advantage of the external information, Authors verified that cooperative nodes having the interlayer links
are much more competent than defectors in sustaining compact clusters of followers. Such an asymmetric exploitation
confers them a strong evolutionary advantage that may resolve even the toughest of social dilemmas.
6. Synchronization
Networked systems, i.e. ensembles of distributed dynamical systems that interact over complex wirings of connections,
are prominent candidates to study the emergence of collective organization in many areas of science [2,9].
′
In particular, a special interest has been devoted to the analysis of synchronized states (from the Greek συγχρoνoς,
meaning ‘‘sharing the same time’’), as such states play a crucial role in many natural phenomena, like as the emergence
of coherent global behaviors in both normal and abnormal brain functions [365], or the food web dynamics in ecological
systems [366].
In the past years, the emergence of synchronized states [367] has been extensively reported and studied for the case of
mono-layer networks, with the emphasis focusing on how the complexity of the layer topology influences the propensity
of the coupled units to synchronize [368–371]. While the literature on this subject is indeed extremely large, we point the
reader to a few monographic reviews and books [9,11,372–374], whose consultation may provide guidance and help in
understanding the major accomplishments, results and related concepts.
As soon as the multilayer nature of connections is brought into the game, the subject must be divided into two different
scenarios, which have to be individually treated.
In the first case, the different layers correspond to different connectivity configurations that alternate to define a time-
dependent structure of coupling amongst a given set of dynamical units. In this scenario, the basic problem is to describe
and assess how the existence and stability conditions for the state in which the network’s dynamical units are synchronous
are modified with respect to the classic (monolayer) situation, in which the coupling between the networked units is time-
independent.
The second case, instead, is the scenario where the different layers are simultaneously responsible for the coupling of the
network’s units, with explicit additional layer–layer interactions to be taken into account. It is worth stressing that here
an interlayer synchronized state is a different and much more general concept, as it does not necessarily require intralayer
synchronization. In other words, the dynamical units within each layer can be out of synchrony, but synchronized with the
corresponding units in all other layers.
As a consequence, this section is divided into two main parts, each reporting on one of the two described scenarios, and is
intended to summarize for each case the extension of the main analytical treatments and approaches that are traditionally
used to deal with synchronization of monolayer networked systems.
6.1. Synchronization in multilayer networks with alternating layers
We start with the case in which the network under consideration is composed of a number N of identical dynamical
systems, which interact via a time-dependent structure of connections, each of which corresponding, at a given time, to one
of the different layers.
Very recent investigations on synchronization of time-varying networks include the treatment of complete
synchronization in temporal Boolean networks [375], the analysis of stability of synchronized states in switching networks
in the presence of time delays [376], and the investigation of the effects of temporal networking interactions when compared
with the corresponding aggregate dynamics where all interactions are present permanently [377].
We focus here on the situation of an ensemble of identical units where the wiring of connections is switching between the
different layers following a periodic or aperiodic sequence. It is important to remark that detailed studies of the latter case
actually started much before the relatively recent introduction of the concept of multilayer networks. In particular, a widely
studied case is the limit of the so-called blinking networks [378,379], where the switching between different configurations
happens rapidly, i.e. with a characteristic time scale much shorter than that of the networked system’s dynamics. Under
these conditions, it was found that synchronous motion can be established for sufficiently rapid switching times even in the
case where each visited wiring configuration (each network layer) would prevent synchronization under static conditions.
Yet, the limit of blinking networks is not a suitable framework for an adequate description of several natural phenomena.
For instance, modeling processes such as mutations in biological systems [380], or adaptations in social systems or financial
market dynamics [381] would demand the very opposite limit of time-varying networks whose structure evolution takes
place over characteristic time scales that are commensurate with, or even secular with respect to, those of the nodes’
dynamics.
Thus, in the following, we will concentrate on the case in which synchronized states appear in dynamical networks,
without making any explicit a priori assumption on the time scale of the variation of the coupling wiring.
6.1.1. The Master Stability Function for time dependent networks

The starting point is the generic equation describing the evolution of an ensemble of N identical systems, networking via
a time dependent structure of connections:
N

ẋi = f(xi ) − σ Gij (t )h[xj ], i = 1, . . . , N . (232)
j =1
Here, dots denote temporal derivative, xi ∈ Rm is the m-dimensional vector describing the state of the ith network unit,
and f(x) : Rm → Rm and h[x] : Rm → Rm are two vectorial functions describing the local dynamics of the nodes and the
output from a node to another, respectively.
Furthermore, Gij (t ) ∈ R are the time varying elements of an N × N symmetric matrix specifying the evolution, in strength
and topology, of  the network connections. The matrix G(t ) satisfies the defining properties of a Lagrangian matrix: it has
zero row-sums ( j Gij (t ) = 0 ∀i and ∀t), strictly positive diagonal terms (Gii (t ) > 0 ∀i and ∀t) and strictly non-positive
off-diagonal terms (Gij (t ) ≤ 0 ∀i ̸= j and ∀t).
Our goal is to assess the stability of the synchronized solution x1 (t ) = x2 (t ) = · · · = xN (t ) ≡ xs (t ), whose existence is
indeed warranted by the zero row-sum property of G(t ).
To this purpose, we first notice that G(t ) is symmetric, and therefore it is always diagonalizable and, at all times, admits
a set λi (t ) of real eigenvalues and a set vi (t ) of associated orthonormal eigenvectors, such that G(t )vi (t ) = λi (t )vi (t ) and
vTj · vi = δij .
The zero row-sum condition and the Gershgorin’s circle theorem [382] further ensure that:
1. the spectrum of eigenvalues is such that λi (t ) ≥ 0 ∀i and ∀t;

2. the smallest eigenvalue λ1 (t ) ≡ 0, with associated eigenvector v1 (t ) = √1 {1, 1, . . . , 1}T , entirely defines the
N
synchronization manifold (xi (t ) = xs (t ) ∀i);
3. all the other eigenvalues λi (t ) (i = 2, . . . , N) have associated eigenvectors vi (t ) that form a basis of the space transverse
to the manifold of xs (t ) in the m × N-dimensional phase space of Eq. (232).
Notice that all the eigenvalues except the first are strictly positive for connected graphs (λi (t ) > 0 for i = 2, . . . , N).
Consider then δ xi (t ) = xi (t ) − xs (t ) = (δ xi,1 (t ), . . . , δ xi,m (t )), i.e. the deviation of the ith vector state from the
synchronized state, and construct the N × m column vectors X = (xT1 , xT2 , . . . , xTN ) and δ X = (δ xT1 , . . . , δ xTN ). It is
straightforward to show that the linearization of Eq. (232) around the synchronized solution leads, in linear order of δ X,
to
δ Ẋ = [IN ⊗ Jf(xs ) − σ G(t ) ⊗ Jh(xs )] δ X, (233)
where ⊗ denotes the direct product, and J is the Jacobian operator.

The arbitrary state δ X, at each time, can be written as a sum of direct products of the eigenvectors of G(t ) and a time-
dependent set of row-vectors ηi (t ), i.e.
N

δX = vi (t ) ⊗ ηi (t ),
i =1
with ηi (t ) = (ηi,1 , . . . , ηi,m ). Therefore, applying vTj to the left side of each term in Eq. (233), one finally obtains
N
d  d
ηj = Kj ηj − vTj (t ) · vi (t )ηi , (234)
dt i=1
dt
where j = 1, . . . , N and Kj = Jf(xs ) − σ λj (t )Jh(xs ) .

 
The key point here is noticing that the boxed term in Eq. (234) is what differentiates them from the classical equations
used to calculate the Master Stability Function (MSF) [370], as it explicitly accounts for the time variation of the eigenvector
basis. In the following, we will first discuss the case where such an extra term vanishes, and later the general case of
constructing a quasi-static evolution among different network layers, deriving a proper analytical expression for it.
6.1.2. Commutative layers

Eq. (234) transform into a set of N variational equations of the form
d
ηj = Kj ηj (235)
dt
vTj (t ) · dt vi (t )ηi = 0, i.e. if all eigenvectors are constant in time.

N d
if i=1
This condition can be satisfied in two different ways, namely either the coupling matrix G(t ) is constant, as in the case
of a monolayer network originally considered in Ref. [370], or the temporal evolution occurs through commutative layers.
The latter case means that starting from an initial wiring G0 ≡ G(t = 0), the coupling matrix G(t ) must commute with G0 at
all times t (G0 G(t ) = G(t )G0 ∀t). This was explicitly discussed in Ref. [383], where it was shown that:
1. an evolution along commutative graphs (i.e., different commutative layers of a multilayer network) can be constructed
by starting from any initial wiring G0 , and
2. such an evolution provides a much better condition for the stability of the synchronous state than that given by a
monolayer network.
To show this, the first step is constructing explicitly the set of commutative layers through which the evolution takes
place. To this purpose, one can notice that any initial symmetric coupling matrix can be written as G0 = V Λ0 V T , where
V = {v1 , . . . , vN } is an orthogonal matrix whose columns are the eigenvectors of G0 , and Λ0 = diag(0, λ2 (0), . . . , λN (0))
is the diagonal matrix formed by the eigenvalues of G0 .
At any other time t, a zero row-sum symmetric commuting matrix G(t ), defining the layer visited by the structure
evolution at that time, can be constructed as
G(t ) = V Λ(t )V T , (236)
where Λ(t ) = diag(0, λ2 (t ), . . . , λN (t )) with λi (t ) > 0 ∀i > 1. It is easy to show that G(t ) is positive semidefinite, and
since the canonical basis vectors are not collinear with the eigenvector v1 we have Gii > 0 ∀i. Thus, Eq. (236) provides a
way to generate all possible layers that commute with G(0), are positive semidefinite with zero row-sum, and dissipatively
couple each network unit. Henceforth, we refer to this set of matrices as the dissipative commuting set (DCS) of G(0). It
is important to notice that while not all members of the DCS are true Lagrangian matrices (it is in fact possible for their
off-diagonal components to be positive), nevertheless all commuting Lagrangian matrices are contained in the DCS.
In Ref. [383], the Authors proposed a general method to select a set of commutative matrices within the DCS. The method
is based on starting from a given layer graph G0 , and producing a large set of different realizations of the same graph. This
allows the estimation of the initial distribution p(λ0 ) of the non-vanishing eigenvalues of the set. Then, one can directly
construct Λ(t ) to use in Eq. (236) by randomly drawing a set of eigenvalues either according to p(λ0 ), or uniformly between
λ2 (0) and λN (0). The former approach can be implemented either by using the spectrum of a different realization of G0 (the
so-called eigenvalue surrogate method [383]), or by extracting directly the eigenvalues from the distribution p(λ0 ). The
latter method, instead, simply corresponds to the random selection of the eigenvalues from a uniform distribution. In either
case, the random extractions can be performed in an ordered (0 ≤ λ2 (t ) ≤ · · · ≤ λN (t )) or unordered way, and one only
needs to extract N − 1 eigenvalues, since λ1 (t ) must be always 0.
In addition, Ref. [383] also discusses how the different choices for the construction of commutative layers lead to
modifications of the main topological structures. In general, the resulting matrix G(t ) is a dense matrix that can be associated
to a undirected weighted network, whose weight matrix W (t ) has elements Wii (t ) = 0, and Wij (t ) = |Gij (t )| for i ̸= j. The
main properties that were studied in Ref. [383] are the average value of the strength distribution, the average local clustering
coefficient, and the average shortest path length, defined as follows [2,9,91,384].
For weighted networks, the distance between two adjacent nodes i and j is lij = W1 , and the distance along a path
ij
{n(1), n(2), . . . , n(m)} can be expressed as Ln(1)→n(m) =

m−1
ln(k)n(k+1) . Then, the distance between two non-adjacent
k=1
nodes i and j is the length of one of the shortest paths connecting them:
ℓij = min Li→j .

allpaths i→j
With these definitions, a network’s average shortest path length is [2,9,91,384]

2 
⟨ℓ⟩ = ℓij .
N (N − 1) i,j
i̸=j
Similarly, the clustering coefficient of node i can be defined as

2 
ci = Wij Wim Wjm ,
si (N − 2) j,m
1
 
where si = j Wij is the strength of node i. The average local clustering coefficient ⟨C ⟩ is then given by ⟨C ⟩ = N i ci , while
the strength distribution P (s) characterizes the heterogeneity of the network [91,384].
Table 5 summarizes the behavior of these quantities for the different eigenvalue selection strategies described above. The
initial layer G0 is the Laplacian matrix of a scale-free (SF) network grown using the preferential attachment algorithm [385],
and whose strength distribution follows a power-law P (s) ∼ s−γ . Fig. 47 illustrates the strength distributions of the
network layers constructed with different methods. It is evident that the eigenvalue surrogate method, as well as the ordered
selection of eigenvalues from the initial distribution p(λ0 ), do not change the SF character of the strength distribution (thick
red and dashed blue lines in panel a). At the same time, they yield values for ⟨s⟩ = {s} sP (s), ⟨C ⟩ and ⟨ℓ⟩ that deviate
only slightly from those of the original graph (see Table 5). This indicates that both methods are convenient strategies to
construct a commuting set of layers through which the evolution of the graph substantially preserves all the main topological
features of the initial condition. Conversely, choosing the eigenvalues from a uniform distribution, either in an ordered or in a
unordered way, produces layers that completely destroy the original SF strength distribution (thick red solid line and dashed
blue line in panel b). These correspond to structural configurations for which each measured quantity wildly differs from its
value in G0 . Finally, the intermediate solution of choosing the eigenvalue spectrum from the initial p(λ0 ) in a unordered way,
while still preserving the power-law tail of the strength distribution (solid black line in panel b), still changes substantially
the topological structure of the underlying network.
Table 5
Statistical properties of the commuting layers. The table reports the average
strength ⟨s⟩, the average local clustering coefficient ⟨C ⟩ and the average shortest
path length ⟨ℓ⟩ for the different methods of selecting the set of commuting layers.
In all cases the results are ensemble averages over 100 different realizations of
networks of size N = 500.
Source: Table reproduced figure with permission from Ref. [383].
⟨s⟩ ⟨C ⟩ ⟨ℓ⟩
Initial condition G0 (scale-free) 9.94 5.5773 × 10 −4
5.5144
Eigenvalue surrogate method 11.6881 8.4730 × 10−4 5.2758
Choosing from p(λ0 ) (ordered) 12.7042 9.1235 × 10−4 5.2325
Choosing from p(λ0 ) (unordered) 142.8746 0.0543 3.7144
Uniform distribution (ordered) 123.6484 0.0634 0.9570
Uniform distribution (unordered) 521.7568 0.6125 1.1098
Fig. 47. (Color online) Strength distributions P (s) of the commuting layers. (a) Initial condition G0 (solid black line), G(t ) constructed by the eigenvalue
surrogate method (thick solid red line), and Λ(t ) obtained by randomly choosing the eigenvalues from p(λ0 ) in an ordered way (dashed blue line). (b)
Eigenvalues randomly chosen from p(λ0 ) in a unordered way (solid black line), eigenvalues chosen from a uniform distribution in an ordered way (thick
solid red line), and in a unordered way (dashed blue line). The details of the statistical ensembles are the same as for Table 5.
6.1.3. Enhancing synchronization via evolution through commutative layers

Besides the analysis of changes in the structural properties of the layers belonging to the DCS and generated via different
methods, one also needs to study the consequences of the evolution through commutative layers on the stability of the
synchronization solution.
Because of the commutativity of the layers, Eq. (234) takes the form given by Eq. (235).
Replacing σ λj (t ) by ν in the kernel Kj , the problem of stability of the synchronization manifold becomes equivalent to
studying the m-dimensional parametric variational equation
η̇ = Kν η,
where Kν = [Jf(xs ) − ν Jh(xs )]. This allows to plot the curve Λmax vs. ν , where Λmax is the largest of the m associated
conditional Lyapunov exponents. This curve is classically called the Master Stability Function [370].
For G(t ) = constant, the synchronized state is transversely stable if all λi (i = 2, . . . , N), multiplied by the same coupling
strength σ , fall in the range where Λmax (ν) < 0. When G(t ) evolves inside the DCS, the eigenvectors are fixed in time,
and one can make the hypothesis that at time T = kdt, the modulus of each eigenmode ηi (i ̸= 1) is bounded by the
corresponding maximum conditional Lyapunov exponent
|ηi (T )| k−1
≤ exp (Λmax (σ λi (ndt)) dt) .
|ηi (0)| n=0
It immediately follows that the transverse stability condition for the synchronization manifold is now that ∀i ̸= 1
 T
1
Λmax σ λi t ′ dt′ < 0 .
  
Si = lim (237)
T →∞ T 0
Fig. 48. Time evolution of synchronization error ⟨δ⟩ for τ = 0.2 (empty circles), 0.3 (filled circles), 0.4 (empty squares), 0.5 (filled squares), 0.6 (empty
diamonds), 0.7 (filled diamonds), 0.9 (empty up triangles), 1.1 (filled up triangles), and 1.5 (empty down triangles). In all cases, the results are ensemble
averages over 5 different random initial conditions for a dynamical network of N = 200 identical chaotic Rössler oscillators. Inset: Synchronization time
Tsync vs. switching time τ (open circles refer to different initial conditions, solid line is the ensemble average).
It is extremely important to notice that condition (237) does not require Λmax (σ λi (t )) to be negative at all times.
In fact, one can construct a commutative evolution such that at each time there exists at least one eigenvalue λi (t ) for
which Λmax (σ λi (t )) > 0, and yet obtain a transversely stable synchronization manifold.
The very relevant conclusion is that a multilayer network where different layers are visited at different times may produce
a stable synchronized motion even when the emergence of a synchronous behavior of the units would be prevented by using
any of the layers as a constant wiring configuration. Similar conclusions were independently reached in Ref. [386] for fast
switching times between commutative layers.
A simple example is a periodic evolution between network layers, with period Tp = (N − 1)τ , during which G(t ) is given
by
N −1

G(t ) = Gl χ[(l−1)τ ,lτ ) , (238)
l =1
where χ[(l−1)τ ,lτ ) denotes the characteristic function of the interval [(l − 1) τ , lτ ). Here, the matrices Gl defining the layers
are constructed, as before, as Gl = V1 Λl V1T . Given Λ1 = diag(0, λ2 , . . . , λN ), the set of the other N − 2 matrices Λl
is constructed by subsequently shifting the sequence of the eigenvalues of G1 , i.e. Λ2 = diag(0, λ3 , λ4 , . . . , λN , λ2 ),
Λ3 = diag(0, λ4 , λ5 , . . . , λ2 , λ3 ), etc.
If, for instance, Λmax (σ λ2 ) > 0, then all the layers will have a direction in the phase space along which the
synchronization manifold is transversely unstable. However, if j=2 Λmax σ λj < 0, then condition (237) will be satisfied
 N −1  
in all directions transverse to the synchronization manifold, and therefore the synchronized solution will be transversely
stable.
As an example of such an extreme situation, Fig. 48 shows the results of an evolution of a network where the initial layer
is a SF graph of N = 200 Rössler chaotic oscillators, each obeying Eq. (232) with
x ≡ (x, y, z ), f(x) = (−y − z , x + 0.165y, 0.2 + z (x − 10)) ,

h[x] = y.
The evolution of the wiring follows Eq. (238). For σ = 0.03 one finds that j=2 Λmax σ λj ≈ −9.158, but Λmax σ λj is
N −1    
positive for the first 80 eigenvalues, i.e., the synchronization manifold would be unstable in at least 80 different transverse
directions in each of the layers, if taken as fixed. However, defining the synchronization error as
N
 |xi − x1 | + |yi − y1 | + |zi − z1 |
⟨δ⟩(t ) = ,
j =2
3(N − 1)
and the synchronization transient Tsync as the time needed for ⟨δ⟩(t ) to become smaller than 0.1, we see from the data of
Fig. 48 that the multilayer network is always able to synchronize, with a Tsync that scales almost exponentially with τ .
It is important to remark that, while this numerical example is used here just to illustrate the enhancement of
synchronization in multilayer commuting graphs under extreme conditions, the validity of condition (237) is general, and
by no means limited to periodic evolutions of the connectivity matrices.
A further point that should be stressed is that the multilayer commuting case is substantially different from the fast
switching procedure described in Ref. [378]. Indeed, the stability condition does not impose in principle any limitation on
the switching time τ , making it possible to enhance the collective functioning of the network even for wirings evolving over
secular (yet finite) time scales.
6.1.4. Non-commutative layers

So far, we considered the case of an evolution through commutative layers, clearly demonstrating the importance
of explicitly accounting for a multilayer nature of interactions in networked dynamical systems, and the fact that such
multilayer structure can greatly enhance the stability of the synchronization solution. Yet, it is evident that considering
commutative layers is too strong an assumption to adequately represent the general case encountered in real world
multilayer graphs.
When explicitly dealing with non commutative layers, the boxed term in Eq. (234) no longer vanishes. Thus, Eq. (234) do
not reduce anymore to a set of variational equations, and one must examine in detail the effects of this extra term on the
coupling of the different directions of stability for the synchronous solution. To do so, there are some issues that need to be
properly dealt with.
The first is that the boxed term in Eq. (234) is explicitly proportional to the time derivatives of the eigenvectors of
the network Laplacian, and therefore any evolution involving instantaneous jumps from a layer to another of a multilayer
structure would result in a divergence. As a consequence, one must consider a quasi-static evolution that depends on two
different time scales: first, the network structure is kept constant from time t = 0 to t = t0 , period over which it is described
by one given layer of the network; then, from t = t0 to t = t1 , a smooth switching process takes place, bringing the network
from one layer to another.
Second, the switching process should be carefully constructed. Indeed, one needs to find a smooth evolution that, starting
from a generic layer whose Laplacian is L0 , leads to another generic layer whose Laplacian is L1 , through a set of intermediate
configurations that all maintain the basic conditions for the existence of the synchronized solution. In practice, this means
that the switching process should occur through a path of connecting matrices, each displaying the zero-row sum condition,
each being diagonalizable, and each sharing the same eigenvector corresponding to the null eigenvalue.
To achieve this, we first consider the matrices A and B, consisting of the eigenvectors of L0 and L1 , respectively, and
describe how to transform one into the other by means of a proper rotation, and in particular a proper rotation around a
fixed axis. Then, we use this framework to find a transformation between L0 and L1 .
Eigenvector matrices. Let A and B be two N × N orthogonal matrices. In our case, we know that the vectors of A and B
correspond to the eigenvectors of two network Laplacians. We also know that at least one vector a is common between A
and B. This vector a corresponds to the null eigenvalue of the Laplacian, and thus to the synchronization manifold.
We want to find a one-parameter transformation group Gs such that:
• G0 A = A;
• Gs A contains the vector a for all 0 6 s 6 1;
• G1 A = B.
Existence of a solution. To show that it is always possible to find a solution, first notice that, in general, the transformation
O from A to B is a rotation, because the vectors of both matrices form orthonormal bases of the same space. To find O, one
can simply solve the linear system of N 2 equations in N 2 variables OA = B. As we are free of choosing the set of coordinates,
we put ourselves in the basis defined by the matrix A. In this basis, A ≡ 1, where 1 is the identity.
Now, without loss of generality, assume that the conserved vector a is the first vector of A, (1, 0, . . . , 0)T . Then, the
transformation matrix O will have the form
1 0 0 ··· 0
 
0 O22 O23 ··· O2N 
0
O= O32 O33 ··· O3N 
. (239)
 ... .. .. .. .. 

. . . . 
0 ON2 ON3 ··· ONN
Notice that, O is a unitary matrix, and therefore the absolute value of |O| is equal to 1. In general O is an element of the
orthogonal group O (N ), which means that it is a proper rotation, if its determinant is 1, or a composition of rotations and
reflections, if its determinant is −1. However, as the form of O in Eq. (239) indicates, its determinant equals the determinant
of the minor O′ obtained by removing from O the first row and the first column. Then, we only need to find a solution to the
problem in N − 1 dimensions. We will henceforth use primes when referring to objects in N − 1 dimensions.
From these considerations, we have G′0 = 1′ and G′1 = O′ . But O′ itself is trivially an orthogonal matrix, belonging to
O (N − 1). Thus, our problem is equivalent to determining the possibility of finding a path between the identity and O′ .
At this point,
  we have two cases.
Case 1: O′  = 1. If the determinant of O′ is positive, then O′ belongs to the special orthogonal group SO (N − 1). In
this case, describing a solution is straightforward. In fact, SO (N − 1) is the identity component of the orthogonal group.
Fig. 49. Example of label swapping. Going from the basis on the left to the one in the middle while keeping x constant necessarily involves a reflection.
However, going from the basis on the right to the one in the middle can be accomplished via a proper rotation.
Moreover, as the orthogonal group is a Lie group, SO (N − 1) is not only connected, but also path-connected. This means
that for every orthogonal (N − 1) × (N − 1) matrix, there is a continuum of orthogonal matrices of the same dimension
connecting it to the identity. Each point along this path corresponds to an orthogonal matrix that can be embedded in SO (N )
by adding a 1 in the top left corner. Since every embedded matrix along the continuum will keep the vector corresponding
to the synchronization manifold invariant, a parametrization of the path from the identity to O will provide a direct solution
to the original
  problem.
Case 2: O′  = −1. If the determinant of O′ is negative, then O′ belongs to O (N − 1)\ SO (N − 1). The same considerations
as for Case 1 apply. However, while O (N − 1)\ SO (N − 1) is also a connected topological space, the identity does not belong
to it. This means that there is no path that can connect the identity to O′ , and we cannot solve the problem the same way.
Yet, in our particular case, we do not really care about the labeling of the vectors. In other words, provided that we
keep the vector a constant, we can very well swap two vectors in the basis given by the matrix A. This literally just means
creating a new matrix C by swapping two rows in A. Of course, this imposes a swap of the corresponding columns in the
transformation matrix O as well. But we also know that swapping two rows (or columns) in a matrix changes the sign of its
determinant. This means that the new matrix O′ does not belong to O (N − 1) \ SO (N − 1), but rather to SO (N − 1). Thus,
it is path-connected to the identity, and all the procedure of Case 1 can be applied. The only consequence of the swap is that
the implicit order of the eigenvalues will be changed. As this order is irrelevant for our purposes, the problem can always
be solved. To illustrate this, consider the following simple example in R3 .
Suppose we want to go from the basis on the left in Fig. 49 (the canonical basis of R3 ) to the one in the middle. Then, the
transformation matrix is
1 0 0
 
Ō = 0 1 0 .
0 0 −1
The determinant of Ō is −1, so there is no continuous way to go from the identity to Ō. However, suppose we relabel the y
and z vectors in our original basis, obtaining the basis in the right panel of Fig. 49. Then, the transformation matrix will be
1 0 0
 
O= 0 0 1 .
0 −1 0
This time, its determinant is 1. Thus O ∈ SO (3), or, equivalently, O′ ∈ SO (2). This means we can find a path that connects
the identity to O. This path can be readily turned into a family of transformations:
1 0 π  0 π 
 
0 cos − s − sin − s 
Gs = 
 π2   π2   .


0 sin − s cos − s
2 2
Note that for every value of the parameter s between 0 and 1, the determinant of Gs equals 1. Also, for any value of s, the
vector x remains always constant. However, the main point here is that while Ō leaves y unchanged and reflects z across the
origin, O instead rotates the ‘‘new y’’ onto the position previously occupied by y, and the ‘‘new z’’ onto the position opposite
to that previously occupied by z. In other words, the final arrangement of the original transformation is preserved, but this
time the transformation is a proper rotation. As we are only interested in getting the final result right, and not in the order
of the eigenvectors, we can disregard this difference.
6.1.5. Building a solution for a generic layer switching

Given a specific A and B, cast in the form discussed above, we build an explicit solution by factoring the transformation
O′ into rotations and reflections in mutually orthogonal subspaces. Before describing the actual procedure, we recall two
useful results. Proving these results involves constructions we will use in the solution and provides further insight into our
original problem. Thus, we present the proofs in the following.
Existence of invariant subspaces. First, any orthogonal operator X in a normed space over R induces an invariant subspace.
This subspace is either 1-dimensional or 2-dimensional. To see this, given X , define the operator U that acts on a + ib as
U (a + ib) = Xa + iXb. As U is a unitary operator, its spectrum lies on the unit circle. Then, one can find a non-vanishing
eigenvector x with eigenvalue λ. In general, x = xR + ixI , and λ = eiϑ = λR + iλI . Then, we have two cases.
Case 1: λ is real. In this case, Ux = XxR + iXxI = ±xR ± ixI . Then, XxR = ±xR and XxI = ±xI . As x ̸= 0,
⃗ at least one between
xR and xI cannot vanish. Without loss of generality, let the non vanishing component be xR . Then, all vectors proportional
to xR are mapped by X into vectors proportional to xR . Thus, the 1-dimensional subspace of the vectors proportional to xR is
invariant under X .
Case 2: λ is complex. In this case, xR and xI are linearly independent. We prove this by contradiction. First note that
x̄ ≡ xR − ixI is an eigenvector of U with eigenvalue λ̄ = e−iϑ = λR − iλI , and that xR = x − ixI = x̄ + ixI .
Now assume that it were xR = α xI . Then one could write
α UxI = UxR = U (x − ixI ) = λx − iUxI ,
hence
(α + i) UxI = (α + i) XxI = λx = λ (α xI + ixI ) = (α + i) λxI ,
from which it would follow that
XxI = λxI . (240)
Similarly, one could write
α UxI = UxR = U (x̄ + ixI ) = λ̄x̄ + iUxI ,
hence
(α − i) UxI = (α − i) XxI = λ̄x̄ = λ̄ (α xI − ixI ) = (α − i) λ̄xI ,
from which it would follow that
XxI = λ̄xI . (241)
Thus, Eqs. (240) and (241) would imply that λ = λ̄. But this directly contradicts the hypothesis that λ ̸∈ R.
Having proved that xR and xI are linearly independent, apply U to x:
Ux = XxR + iXxI = λx = λR xR − λI xI + i (λI xR + λR xI ) ,
hence
XxR = λR xR − λI xI (242)
and
XxI = λI xR + λR xI . (243)
The two equations above imply that
X (α xR + β xI ) = αλR xR − αλI xI + βλI xR + βλR xI = (αλR + βλI ) xR + (βλR − αλI ) xI = γ xR + δ xI ,
where γ = αλR + βλI and δ = βλR − αλI . This means that any linear combination of xR and xI is mapped by X into a linear
combination of the same vectors xR and xI . As we have proved above that they are linearly independent, this implies that
the 2-dimensional subspace determined by the basis {xR , xI } is invariant under application of X .
Orthogonality of the eigenvector components. We show now that the components xR and xI of the eigenvector x used in Case 2
of the previous proof are not only linearly independent, but actually orthogonal. To see this, compute the inner product
between xR and xI , exploiting the fact that the operator X preserves the inner product, since it is orthogonal:
⟨xR , xI ⟩ = ⟨XxR , XxI ⟩ = ⟨λR xR − λI xI , λI xR + λR xI ⟩ , (244)
where we have used Eq. (242) and (243).
Then it is
⟨xR , xI ⟩ = λR λI ∥xR ∥2 + λ2R ⟨xR , xI ⟩ − λ2I ⟨xI , xR ⟩ − λR λI ∥xI ∥2 = λR λI ∥xR ∥2 − ∥xI ∥2 + λ2R − λ2I ⟨xR , xI ⟩ ,
   
hence
λR λI
⟨xR , xI ⟩ = ∥xR ∥2 − ∥xI ∥2
 
1 − λ2R + λ2I
λR λI 
∥x R ∥2 − ∥x I ∥2

= (245)
2λI2
λR 
∥xR ∥2 − ∥xI ∥2 ,

=
2λI
where we have used λ2R + λ2I = 1.
Similarly, applying the same property of X , and using Eq. (242), the squared norm of xR is
∥xR ∥2 = ⟨xR , xR ⟩
= ⟨λR xR − λI xI , λR xR − λI xI ⟩
= λ2R ∥xR ∥2 − λR λI ⟨xR , xI ⟩ − λR λI ⟨xI , xR ⟩ + λ2I ∥xI ∥2
= λ2R ∥xR ∥2 − 2λR λI ⟨xR , xI ⟩ + λ2I ∥xI ∥2
= λ2R ∥xR ∥2 + 1 − λ2R ∥xI ∥2 − 2λR λI ⟨xR , xI ⟩
 
= λ2R ∥xR ∥2 − ∥xI ∥2 + ∥xI ∥2 − 2λR λI ⟨xR , xI ⟩ .

 
Then, it is
∥xR ∥2 − ∥xI ∥2 = λ2R ∥xR ∥2 − ∥xI ∥2 − 2λR λI ⟨xR , xI ⟩ ,

 
and, using again λ2R + λ2I = 1,
2λR
∥x R ∥2 − ∥x I ∥2 = − ⟨xR , xI ⟩ . (246)
λI
Substituting Eq. (246) into Eq. (245) yields
λ2R
⟨xR , xI ⟩ = − ⟨xR , xI ⟩ .
λ2I
Now, we know by hypothesis that λI ̸= 0. Then, if λR ̸= 0 it must be ⟨xR , xI ⟩ = 0. Alternatively, if λR = 0, then it is
λI = ±1. In this case, from Eq. (244), it follows that
⟨xR , xI ⟩ = ⟨∓xI , ±xR ⟩ = − ⟨xI , xR ⟩ ⇒ ⟨xR , xI ⟩ = 0.
Either way, xR and xI are orthogonal.
The solution. Using the two results just shown, we can describe an algorithmic procedure to find the solution to our problem.
The initial situation is that we have two orthonormal bases A and B of RN . Then, we can build a transformation of A into B
with the following steps:
1. Express the problem in the basis A. In this basis, the transformation matrix O between A = 1 and B takes the form given
by Eq. (239).
2. Consider the operator O′ obtained from O by removing the first row and the first column. Let d be its dimension.
3. Build the operator U that acts on a + ib as U (a + ib) = O′ a + iO′ b, where a and b belong to Rd .
4. Find an eigenvector x = xR + ixI of U, with eigenvalue λ.
5. Normalize xR and xI .
6. If λ ∈ R
(a) Pick the non-vanishing component between xR and xI . If both are non-zero, choose one of them randomly. Without
loss of generality, assume this is xR .
(b) Create d − 1 other orthonormal vectors, all orthogonal to xR , and arrange all these vectors so that xR is the last of
them. This set of vectors is an orthonormal basis C of Rd .
(c) Change the basis of the d-dimensional sub-problem to C . In this basis, all the elements in the last row and in the last
column of O′ will be 0, except the last one, which will be ±1.
(d) If d > 1, consider a new operator O′ obtained from the old O′ by removing the last row and the last column. Let d be
its dimension, and restart from step 3. Otherwise stop.
7. If λ ̸∈ R
(a) Create d − 2 other orthonormal vectors, all orthogonal to xR and xI , and arrange all these vectors so that xR and xI are
the first two of them. This set of vectors is an orthonormal basis C of Rd .
(b) Change the basis of the d-dimensional sub-problem to C . In this basis, all the elements in the first two rows and in
the first two columns of O′ will be 0, except the first two.
(c) If d > 2, consider a new operator O′ obtained from the old O′ by removing the first two rows and the first two
columns. Let d be its dimension, and restart from step 3. Otherwise stop.
The steps described above effectively solve the problem of finding the basis transformation by divide-and-conquer. At
each repetition, 1 or 2 dimensions are eliminated from the problem. All the subsequent changes of bases do not affect
the elements of the transformation matrix determined in the previous steps, because they act on subspaces that are
orthogonal to those that have been already eliminated, which include the top left element in A and B, corresponding to the
synchronization manifold. Thus, one can reconstruct the transformation matrix O piece by piece. At the end of the procedure,
O has a block-diagonal form
1 0 0 0 0 ··· 0 0

0 O22 O23 0 0 ··· 0 0 
0
 O32 O33 0 0 ··· 0 0 

0 0 0 O44 O45 ··· 0 0 
O = 0 ,
 
0 0 O54 O55 ··· 0 0
 .. .. .. .. .. .. .. ..
 
. . . . . . . .


0 0 0 0 0 ··· ON −1N −1 ON −1N
 
0 0 0 0 0 ··· ONN −1 ONN
where the blocks correspond to the action of the orthogonal operator on the invariant subspaces. If the invariant subspace
is 1-dimensional, then its block is a single ±1 element. Conversely, if the subspace is 2-dimensional, then its block is either
a rotation or a reflection. This means that it is either
cos α − sin α
 
sin α cos α
or
 
±1 0
.
0 ∓1
At this point, one can proceed as follows. First, permute the basis vectors from the second onwards so that they
correspond, in order, first to all the actual rotation blocks, then to the −1 elements, and finally to the +1 elements. It is
easy to see that this can always be done. In fact, if we call our change-of-basis matrix T , we can determine the new form of
the transformation matrix, ON , via the following logical chain:
OKA = KB ⇒ ON TKA = TKB ⇒ T −1 ON TKA = KB ⇒ T −1 ON T = O ⇒ ON = TOT −1 ,
where K is the change-of-basis matrix corresponding to the composition of all changes of bases done in steps 6 or 7 above.
As the matrix T that swaps the position of the basis vectors is a permutation matrix, it follows that ON has exactly the
required form
1 0 0 ··· 0 0 0 0 0 ··· 0 0
cos ϑ1 − sin ϑ1 ··· ···
0 0 0 0 0 0 0 0
sin ϑ1 cos ϑ1

0 ··· 0 0 0 0 0 ··· 0 0
. .. .. ..

.
. . . .

0 0 0 0 0 ··· 0 0
cos ϑk − sin ϑk
 
0 0 0 0 0 0 0 ··· 0 0
O = 0 0 0 0 sin ϑk cos ϑk 0 0 0 ··· 0 0 . (247)
 
0
 0 0 0 0 0 −1 0 0 ··· 0 0

0 0 0 0 0 0 0 −1 0 ··· 0 0
. .. .. .. .. .. .. .. .. .. .. .. 
 
 .. . . . . . . . . . . .
 
0 0 0 0 0 0 0 0 0 ··· 1 0
0 0 0 0 0 0 0 0 0 ··· 0 1
Having obtained the form in Eq. (247), there is still the possibility that the determinant of O is −1. However, as we
saw, in this case we can approach the problem by relabeling two vectors of the original basis, which induces a swap of the
corresponding columns of O. To perform this, first note that, if |O| = −1, then the number of −1 elements in O must be odd.
Then, there are three possible cases.
Case 1: O′ has at least one +1 element. In this case, permute once more the basis vectors, swapping
 those
 corresponding
1 0
to the first −1 and the first +1 elements in O′ . Then, the first block after the ‘‘sin–cos’’ blocks is 0 −1 . Now, relabel the
two corresponding vectors, as in the R3 example, swapping the relevant rows in the starting matrix. This makes the block
in O′ become
  π  π 
  cos − − sin −
0 1
=  π2   π2  .
 
−1 0
sin − cos −
2 2
Case 2: O′ has no +1 elements and only one −1 element. In this case, the basis vectors to swap are those corresponding
to the first two vectors in the last 3 × 3 block of O′ . This will become
− sin ϑk cos ϑk 0
 
M = cos ϑk sin ϑk 0 .
0 0 −1
Now, change once more the basis of the problem. In particular, leave all the basis vectors unchanged, but map the last
three into the eigenvectors of M. This is easily accomplished, as the matrix of eigenvectors of M is
π ϑk ϑk π
    
0 (− sec ϑk − tan ϑk ) sin − (sec ϑk − tan ϑk ) sin +
4 2 2 4
V = .
 
0 1 1
1 0 0
Then, the new form of M is
−1 0 0
 
′
M = V MV = T
0 −1 0 .
0 0 1
Case 3: O′ has no +1 elements and at least three −1 elements. Similar to the previous case, swap the basis vectors
corresponding to the first two after the ‘‘sin–cos’’ blocks. Their 3 × 3 block is now
0 −1 0
 
M = −1 0 0 .
0 0 −1
Next, do a change of basis as described in Case 2. The eigenvector matrix of M is
1 1
 
0 √ −√ 
 2 2
,

V =
0 1 1 
√ √ 
2 2
 
1 0 0
so after the basis change the new form of M is once more
−1 0 0
 
′
M = V MV = T
0 −1 0 .
0 0 1
Whether the determinant of O was +1 to start with, or it was −1 and taken care of as one of the cases above, there are
now either no −1 elements in O, or an even number of them. Then, if there are any, take every two subsequent −1 elements
and change their diagonal blocks into
cos π − sin π
 
.
sin π cos π
This yields the final general form for the transformation matrix:
1 0 0 ··· 0 0 0 0 0 0 0 ··· 0 0
0 cos ϑ1 − sin ϑ1 ··· 0 0 0 0 0 0 0 ··· 0 0
sin ϑ1 cos ϑ1 ··· ···
0 0 0 0 0 0 0 0 0 0


. .. .. ..

.
. . . .

0 0 0 0 0 0 0 ··· 0 0
cos ϑk − sin ϑk
 
0
 0 0 0 0 0 0 0 0 ··· 0 0

0 0 0 0 sin ϑk cos ϑk 0 0 0 0 0 ··· 0 0
π π
  
 
0 0 0 0 0 0 cos − − sin − 0 0 0 ··· 0 0
O= 2  . (248)
  π  π2 
sin − cos − ···
0 0 0 0 0 0 0 0 0 0 0

2 2

cos π − sin π
 
0 0 0 0 0 0 0 0 0 ··· 0 0
sin π cos π
 
0 0 0 0 0 0 0 0 0 ··· 0 0
. .. .. .. .. .. .. .. .. .. .. .. .. 
 
. ..
. . . . . . . . . . . . . .

 
0 0 0 0 0 0 0 0 0 0 0 ··· 1 0
0 0 0 0 0 0 0 0 0 0 0 ··· 0 1
Finally, introducing a parameter s ∈ [0, 1], we get the required transformation:

1 0 0 ··· 0 0 0 0 0 0 0 ··· 0 0
0 cos (ϑ1 s) − sin (ϑ1 s) ··· 0 0 0 0 0 0 0 ··· 0 0
0 sin (ϑ1 s) cos (ϑ1 s) ··· 0 0 0 0 0 0 0 ··· 0 0
 
. . . ..
 
. . .

. . . .

0 0 0 0 0 0 0 ··· 0 0
cos (ϑk s) − sin (ϑk s)
 
0 0 0 0 0 0 0 0 0 ··· 0 0
sin (ϑk s) cos (ϑk s)
 
0
 0 0 0
0 π  0 π  0 0 0 ··· 0 0

 
0 0 0 0 0 0 cos − s − sin − s 0 0 0 ··· 0 0
Gs =  .
  π2   π2  
0 0 0 0 0 0 sin − s cos − s 0 0 0 ··· 0 0
 
 2 2 
0
 0 0 0 0 0 0 0 cos (π s) − sin (π s) 0 ··· 0 0

sin (π s) cos (π s) ···
0 0 0 0 0 0 0 0 0 0 0


. . . . . . . . . . . .. . .
 
. . . . . . . . . . . . . .
. . . . . . . . . . . . .
0 0 0 0 0 0 0 0 0 0 0 ··· 1 0
 
0 0 0 0 0 0 0 0 0 0 0 ··· 0 1
(249)
Notice that when s = 0, Gs = G0 = 1, and when s = 1, Gs = G1 = O. Also, for every value of s, the determinant
of Gs is always 1, and the first vector is kept constant, which means that Gs always describes a proper rotation around the
axis defined by the eigenvector corresponding to the synchronization manifold. Moreover, note that the first vector has
always been left untouched by all the possible basis transformations. Thus, as s is varied continuously between 0 and 1, the
application of Gs sends A into B smoothly and continuously, as we need.
Laplacian transformation. To describe how to obtain a transformation between the two Laplacians L0 and L1 , let us first
summarize what we have shown.
We started from the matrices A and B, consisting of the eigenvectors of L0 and L1 . These two matrices are orthonormal
bases of RN . We showed that we can always find a series of basis changes casting the problem in a reference frame where
we can explicitly construct a 1-parameter transformation group Gs that transforms A into B with continuity via a rotation
around an axis determined by the eigenvector corresponding to the synchronization manifold.
In formulae, the group Gs is such that G0 = 1 and G1 RA = RB, where R is the change-of-basis matrix resulting from the
composition of all basis changes done in steps 6 and 7 of the method, the permutation of the vectors to obtain O, and the
eventual final adjustment in case of negative determinant. To simplify the formalism, in the following we let B0 ≡ A and
B1 ≡ B.
Then, since B0 and B1 are matrices of eigenvectors, it is
BT0 L0 B0 = D0
BT1 L1 B1 = D1 ,
where D0 and D1 are diagonal matrices whose elements are the eigenvalues of L0 and L1 .
However, as Gs is a group of orthogonal transformation, we can write that
Gs RB0 = RBs ∀0 6 s 6 1,
where Bs is a basis of RN . Multiplying the equation above by RT on the left, we get
RT Gs RB0 = RT RBs = Bs , (250)

T −1
where we have used the fact that R = R .
Now, note that all the basis changes done to build G are between orthonormal bases. Therefore, their composition R is an
isometry, as is any Gs , since they are all proper rigid rotations. Thus, any Bs still defines an orthonormal basis of RN .
Then, we can say that the Laplacian for the parameter s is given by the matrix Ls that solves the equation
BTs Ls Bs = Ds , (251)
where Ds is a diagonal matrix whose elements are the eigenvalues of Ls , and Bs consists of the eigenvectors of Ls . However,
the equation above has two unknowns, namely Ls and Ds .
To solve this last problem we impose a linear evolution of the eigenvalues of the Laplacians. Thus, for all 1 6 i 6 N,
λ(i s) = (1 − s) λ(i 0) + sλ(i 1) , (252)

(0) (1)
where λi and λi are the ith eigenvalues of L0 and L1 , respectively. Then, multiplying Eq. (251) on the right by BTs , it is
BTs Ls Bs BTs = Ds BTs ,

hence
BTs Ls = Ds BTs .
Multiplying this on the left by Bs , we get
Bs BTs Ls = Bs Ds BTs ,
hence
Ls = Bs Ds BTs .
Substituting Eq. (250) into the equation above, we get
T
Ls = RT Gs RB0 Ds RT Gs RB0 = RT Gs RB0 Ds BT0 RT GTs R.

(253)
In Eq. (253), B0 is known; R and Gs have been explicitly built, so they are known as well; Ds is completely determined
by Eq. (252). Thus, Eq. (253) defines the Laplacian for any given value of the parameter 0 6 s 6 1. The evolution of the
eigenvalues follows a simple linear form, imposed by Eq. (252), and the evolution of the eigenvectors is given by Eq. (250).
As a consistency check, it is straightforward to treat the left-hand side of Eq. (253) as an unknown and solve it for either
s = 0 or s = 1, recovering L0 and L1 , respectively. Similarly, it is easy to see that for any s the determinant of Ls is the product
of the eigenvalues defined by Eq. (252).
With this result, we can finally describe the evolution via non-commutative layers. We use the setting introduced in
Section 6.1.4: at time t = 0, the network starts in a layer whose Laplacian is L0 , in which it remains up to time t0 ; then, it
switches to a layer whose Laplacian is L1 , completing the switch at time t1 . What is needed is to compute the boxed term in
Eq. (234) for all times t.
It is clear that, from t = 0 to t = t0 , the term vanishes. To compute its value during the switch, first note that the ith
t −t
eigenvector at time t is the ith column of Bs , where s ≡ t −t0 . But then, using Eq. (250), it is straightforward to see that the
1 0
kth element of the ith eigenvector is
N 
 N 
N
(vi )k = (Bs )ki = RT Gs RB0 ki = Rrk (Gs )rq Rqx (B0 )xi .
 
r =1 q =1 x =1
Notice that in the equation above the only term that depends on time is (Gs )rq , since R is just a change-of-basis matrix,
and B0 is the matrix of eigenvectors of L0 at time t = 0. Therefore, it is
N 
N 
N
d  1 d
(vi )k = Rrk Rqx (B0 )xi (Gs )rq .
dt r =1 q=1 x=1
t1 − t0 ds
This allows us to write a final expression for the extra term that accounts for the time variation of the eigenvectors. In a
fully explicit form, this is
N
 d
− vTj (t ) · vi (t ) ηi
i=1
dt
  
N 
N N 
N 
N N 
N 
N
1    d
=− Rrk (Gs )rq Rqx (B0 )xj Rrk Rqx (B0 )xi (Gs )rq ηi . (254)
t1 − t0 i=1 k=1 r =1 q =1 x =1 r =1 q=1 x=1
ds
However, for all practical purposes, one does not need to use Eq. (254) directly. In fact, considering that most elements
d
of Gs are 0, it is quite simple to compute and store Bs in a symbolic form. Similarly, most of the terms ds (Gs )rq are 0. In fact,
they vanish in all the following cases:
• r = 1 or q = 1 (the eigenvector corresponding to the synchronization manifold, and axis of rotation of the basis of
eigenvectors);
• if the (rq) element is outside the tridiagonal, i.e. if |r − q| > 1;
• if r > 2b + 1 or q > 2b + 1, where b is the number of ‘‘sin–cos’’ blocks in Gq .
Also, for all other cases d
ds
(Gs )rq , is proportional to a sine or a cosine. Thus, we can define Ġs to be the matrix whose (rq)
element is (Gs )rq ; also, we can define B˙s ≡ RT Ġs RB0 . Again, B˙s can be easily computed and stored in a symbolic form.
1 d
t1 −t0 ds
Then, Eq. (254) becomes
N N 
N
 d 
vTj (t ) · vi (t ) ηi = − (Bs )kj B˙s ki ηi .
 
− (255)
i=1
dt i=1 k=1
Once more, Eq. (255) can be computed symbolically, and evaluated at any particular time t, when needed.
6.2. Synchronization in multilayer networks with coexisting layers
A completely different scenario is that of networking dynamical units with coexisting layers. In this case, indeed, the
different layers determine as a whole the overall coupling structure through which the system components interact, and
which in general cannot be simply represented in terms of a single time-independent coupling matrix.
The current state of the art on the matter includes only a few direct studies, pointing however to some important
considerations and results. In the following, we first summarize the available literature following the categorization
given in Section 2, in the hope to point out promising directions to those researchers who, in the near future, will try
to approach such a challenging issue. After that, we concentrate on a specific case, involving a two-layer network in a
master–slave configuration, in view of the relevance of this situation that extends the very concept of synchronization and
provides practical tools to target the dynamics of complex networks toward any desired collective behavior (not necessarily
synchronous).
6.2.1. Hypernetworks
In this framework, one considers a set of N identical dynamical systems linked through the connections of M different
layers, that is, all the connections that correspond to the same type of coupling form a layer, and the systems are connected
by more than one layer.
Refs. [387,388] focused on the necessary and sufficient conditions for the stability of the synchronous solution in a
hypernetwork. The main conclusion is that, in general, it is not possible to reduce the problem to a simpler monolayer
formalism. However, for M = 2, such a reduction is indeed possible in three cases of interest: (i) if the two associated
Laplacians commute; (ii) if one of the two layers is unweighted and fully connected; (iii) if one of the two networks has an
adjacency matrix of the form Aij = aj with i, j = 1, . . . , N. The major result is the definition of a class of networks such
that if one of the two coupling layers satisfies condition (iii), the reduction can be obtained independently of the structural
form of the other network layer. The Authors of Refs. [387,388] also discuss some generalization of their results to the case
of hypernetworks consisting of more than two layers, and present an example of a network of neurons connected by both
electrical gap-junctions and chemical synapses.
6.2.2. Interconnected/interdependent networks and networks of networks

Other studies dealt with the problem of seeking connection strategies between interconnected networks (different layers
of an overall network) to achieve complete synchronization. In this framework, each of the interconnected layers is a network
made of elements of similar or different nature.
Ref. [389] is a combination of experimental, numerical and analytical studies in the case of only two layers. The Authors
find that connector nodes (those at the ends of interlayer links), play a crucial role for the setting of the synchronous behavior
of the whole system. In particular, the main result is that connecting the hubs of both networks is the best strategy to achieve
complete synchronization.
In Ref. [390] the same problem is investigated in the framework of identical layers. A phase transition in the network
synchronizability is described when the number of interlinks increases. The exact location of the transition depends
exclusively on the algebraic connectivity of the graph models, and it is always observed regardless of the interconnected
graphs. Furthermore, by means of computer simulations and analytical methods, Ref. [391] investigates the effects of the
degree distribution in determining mutual synchronization of neural networks made of two identical layers. The main claim
of Ref. [391] is that a SF topology is more effective in inducing synchronization, or in preventing the system from being
synchronized, provided that the interlayer links involve the nodes with larger degrees.
Refs. [392,393] studied synchronization in a complex clustered network (or network of clusters) when two clusters
interact via random connections, showing that the interplay between interconnections and intraconnections plays a
key role in the global synchronizability, as well as in determining different synchronization scenarios. For instance,
Ref. [392] demonstrated that the network has the strongest synchronizability only when the number of inter and
intraconnections matches perfectly, whereas any mismatch weakens or even destroys the synchronous state. However,
Ref. [393] pointed out that in a modular network of phase oscillators there exists an optimal balance condition between
the number and weight of inter and intralinks that maximizes at the same time the network’s segregation and integration
processes.
Furthermore, in networks of interacting populations of phase oscillators, the onset of coherent collective behavior
was studied in Ref. [394]. A typical biological application of the model is the description of interacting populations of
fireflies, where each population inhabits a separate tree. The model includes heterogeneity at several levels. Each population
consists of a collection of phase oscillators whose natural frequencies are drawn at random from a given distribution.
To allow for heterogeneity at the population level, each population has its own frequency distribution. In addition, the
system is heterogeneous at the coupling level, as the model explicitly considers the possibility to specify separately
the coupling strengths between the members of different populations. The assumption of a global (yet population-
weighted) coupling allows for an analytical determination of the critical condition for the onset of a coherent collective
behavior.
6.2.3. Multiplex, multiplex with time delay and master–slave configurations

Ref. [395] reports on the synchronization properties of interconnected networks of oscillators with a time delay on the
interactions between networks, and analyzes the synchronization scenario as a function of the coupling and communication
lag. In particular, the Authors consider the competition between an instantaneous intranetwork and a delayed internetwork
coupling. The major result is the discovery of a new breathing synchronization regime, with the emergence of two groups in
each network synchronized at different frequencies. Each group has a counterpart in the opposite network, one group being
in phase and the other in anti-phase with the counterpart. For strong coupling strengths, however, networks are internally
synchronized, in some cases with a phase shift between them.
In Ref. [396] the reader can find an interesting study on the Master Stability Function for class III systems [9]. The
motivation is to investigate how the topologies of the layers influence the synchronization criteria and rate for the multiplex
network. For that purpose, Ref. [396] considers different topologies, such as random (ER), random regular (RR), small-world
(SW) and scale-free (SF), and studies a 2-layered multiplex network formed by interconnecting a layer with ER, SW or SF
structure with another layer with an RR structure. This choice of the global topology is motivated by its correspondence
to the highest synchronizability. The links between the layers have a constant weight c. The inspection of the eigenratio
λN /λ2 allows to draw non-trivial and interesting conclusions for the synchronization properties. In particular, it is shown
that a SW–RR topology enhances the synchronization, even though individually the SF topology has a lower eigenratio
and a better synchronizability than the SW topology. In addition, it is shown that the ER–RR configuration displays similar
synchronization properties to the SW–RR configuration for small coupling strengths, whereas at large coupling strengths
the ER–RR configuration leads to a synchronization scenario similar to that occurring for the SF–RR configuration.
In Ref. [397], using Lyapunov stability theory, the Authors predict theoretically that in a master–slave configuration,
complete synchronization between master and slave can be achieved if they have identical connection topologies. This
prediction is verified numerically for a two-layer network of Lorenz systems. Ref. [397] also provides some numerical test for
the case in which the master and the slave layers have non-identical connection topologies. In this case, the behavior of the
system is highly non-trivial, with synchronization and de-synchronization alternately appearing for increasing values of the
coupling strength. In this very same context, a fault-tolerant control scheme has been proposed to guarantee synchronization
when the coupling connections fail [398].
Other forms of synchronization, such as generalized synchronization, have been explored in networks with unidirectional
coupling in a master–slave configuration [399–402].
Finally, in Ref. [403] a relatively simple model is studied, consisting of a pair of bidirectional sub-networks each with
an arbitrary number of identical neurons in a ring structure and two-way couplings between the ring sub-networks.
Different time delays are introduced in the internal connections within each layer, and in the couplings between the
layers. The network exhibits different dynamical phenomena depending on the parity of the number of neurons in the
sub-network. Multiple stability switches of the network equilibrium, synchronous and asynchronous periodic oscillations,
and the coexistence of the bifurcated periodic oscillations are shown numerically.
6.2.4. Bipartite and multilayer networks

Concurrent synchronization or multisynchronous evolution in networks is a dynamical regime where multiple groups of
fully synchronized elements coexist in distinct, possibly chaotic, dynamics, so that elements from different groups are not
necessarily synchronized. This kind of behavior has been studied in relation to robotics and neuroscience, where groups of
interacting robots or neurons perform synchronous parallel coordinated tasks.
In Refs. [404,405] the issue of multiple synchronized, possibly chaotic, motions occurring in networks of groups has
been independently addressed. In the first study, the Authors tackled the stability of an invariant subspace corresponding
to multisynchronicity in unweighted networks. In the second, the stability of the multisynchronous evolution in bipartite
networks is evaluated by introducing a MSF that decouples the effects of the network topology from those of the
dynamics on the nodes. The MSF approach seems to be inadequate to assess the stability of the synchronous evolution
when both intragroup (intralayer) and extra-group (interlayer) connections are allowed in the network. The Authors
considered examples of general network topologies, where links are also allowed to fall within each layer, and reported
numerical evidence that the presence of diffusive couplings amongst nodes within the same layer can enhance the network
synchronizability.
Based on Lyapunov stability theory and linear matrix inequality, the Authors of Ref. [406] studied ‘‘inner’’ (within each
network) and ‘‘outer’’ (between two networks) synchronization between two coupled discrete time networks with time
delays. They were able to provide sufficient conditions to assess the stability of the synchronous solution within each
network, but once again the method used cannot guarantee the stability when both intra-group and extra-group connections
are allowed.
In Refs. [407–409] the synchronization of one-layer [407], two-layer [408], and three-layer center networks [409] of
coupled maps has been studied, using linear stability analysis. Multicenter layer networks are characterized by the presence
in each layer of nodes with special properties, called center nodes. Each layer is considered adjacent to the layer immediately
preceding it and the one immediately following it; the exceptions are the first layer, which is only adjacent to the second,
and the last, which is only adjacent to the previous one. All center nodes in a layer are not connected among them (except
for the first layer in which they are globally connected) but to all the center nodes in the adjacent layers. In addition, each
center node in the last layer has n different non-center nodes which are not connected among them. The Authors found
conditions for stable synchronization that depend on the number of center nodes and non-center nodes.
A natural framework for synchronization in multilayer networks is that of transmission of information in neural
networks. A debate on the role of synchrony in nervous system signaling is addressed in Ref. [410], where a detailed analysis
of the relationship between synchronization and rate encoding has been carried out in models of feed-forward multilayer
neuronal networks [410–412]. In Ref. [412], a feed-forward multilayer network of Hodgkin–Huxley neurons was constructed
(with no intralayer links), where synchronous firings can develop gradually within the layered network, consistently with
experimental findings [410] and with other theoretical work [411].
An analogy between crowd synchrony and multilayer neural network architectures is proposed in Ref. [413]. The Authors
showed that a two-layer system of many non-identical oscillators (semiconductor lasers) communicating indirectly via a
few mediators (hubs) can synchronize when the number of optically delayed couplings to the hubs, or the strength of the
coupling, is large enough. The synchronization transition is discontinuous, but its order depends on the detuning between
the hubs. If the detuning is increased, the transition becomes continuous, with a diverging synchronization time near the
critical coupling that does not depend on the number of hubs.
6.2.5. Synchronization and the problem of targeting

A very important point to stress is that, as long as the nodes are identical dynamical systems represented by vector states
of the kind xli (t ), with i = 1, . . . , N indicating the node, and l = 1, . . . , M the layer, the analysis of the synchronization
properties may be one of two very different endeavors.
The first involves an extension of the usual synchronization concept that we call intralayer synchronization, and
corresponds to analyzing the conditions for which the solution limt →∞ |xli (t ) − xlj (t )| = 0 (with i ̸= j and ∀l) is stable.
Notice that the stability conditions of such a solution can be simply found by considering an overall modular-like network
of size NM, and applying to usual MSF formalism [9] to study the spectral properties of the NM × NM Laplacian matrix,
which will have the form of M diagonal blocks describing the intralayer connection structures, plus extra blocks describing
the interlayer interactions.
The second case is much more general, as it encompasses explicitly the first case as one of its particular instances. It
′
corresponds to assessing the stability conditions of the solution limt →∞ |xli (t ) − xli (t )| = 0 with l ̸= l′ and ∀i. Such a
solution, which we call the interlayer synchronized solution, does not imply necessarily that the dynamical evolution of
nodes be synchronized within each layer.
In particular, when restricting the framework to a two-layer network with unidirectional interactions from a master
layer to a slave layer, the importance of the interlayer solution was stressed in connection with the problem of targeting the
dynamics of a complex network [414], i.e. finding a generic procedure to steer the dynamics of a network toward a given,
desired evolution.
In fact, in Ref. [414], the feasibility of targeting any goal dynamics g(t ) compatible with the natural evolution of the
network was proven to be tantamount to demonstrating that two identical layers coupled in a master–slave scheme can
identically synchronize [415,416] when starting from different initial conditions. There, the Authors considered two identical
layers, namely a master network (MN) providing the specific g(t ) and a slave network (SN). Starting from different initial
conditions, both layers produce a turbulent regime, i.e. generic dynamics not showing any particular global or local order. The
Authors properly engineered a unidirectional pinning action from the MN to the SN with the aim of ultimately synchronizing
the dynamics of the two layers.
In the following, we briefly summarize the main results of Ref. [414].
The master network considered is composed of N identical, diffusively coupled chaotic units. Calling the vector state of
the ith node xM m
i ∈ R , one has:
N

i = f(xi ) − σ1
ẋM Lij h1 (xM
j ),
M
(256)
j =1
where f(x) : Rm → Rm and h1 (x) : Rm → Rm are a local evolution and an vectorial output function, respectively. σ1 is a
coupling strength, and Lij are the elements of the corresponding zero row-sum Laplacian matrix of the interaction network.
Notice that, here and in the following, M indicates the master network, and not the number of layers, which is here always 2
(the master and the slave). f and h1 are chosen to make the network units of class I, with a monotonically increasing trend
of the MSF associated to the intralayer synchronous state xM M M
1 = x2 = · · · = xN = xsync [9], which is, therefore, always
unstable, regardless of the choice of σ1 .
Let now
N

ẋSi = f(xSi ) − σ1 Lij h1 (xSj ) (257)
j =1
be the slave network (SN), obtained as a copy of the MN.

As long as both networks remain uncoupled and start from different initial conditions, each one will sustain a different
turbulent state, so that XS (t ) ≡ (xS1 (t ), . . . , xSN (t )) ̸= XM (t ) ≡ (xM
1 (t ), . . . , xN (t )) = g(t ), and the SN will never attain the
M
desired goal dynamics.

In Ref. [414], the Authors implemented a pinning strategy to target g(t ) consisting of sequentially establishing
unidirectional links between nodes in the MN and their copies in the SN. The dynamical evolution of the SN is therefore
described by:
N

ẋSi = f(xSi ) − σ2 χi h2 (xSi − xM
i ) − σ1 Lij h1 (xSj ), (258)
j =1
where h2 : Rm → Rm is the coupling function between MN and SN, with the condition that h2 (0) = 0, and χi encodes
the pinning procedure, i.e. χi = 1 if there is a link from the ith node of the MN to the ith node of the SN, and 0 otherwise.
Furthermore, σ2 is the parameter ruling the interlayer link strength.
It follows that the equation for the vector describing the difference between the master and the slave layers’ dynamics
δ X = XS − XM ≡ (δ x1 , . . . , δ xN ) can be written in terms of its components δ xi as
N

δ ẋi = f(xSi ) − f(xM
i ) − σ1 Lij [h1 (xSj ) − h1 (xM
j )] − σ2 χi h2 (xi − xi ).
S M
(259)
j =1
A stable fixed point of Eqs. (259) at δ X = 0 is a necessary condition for layers (256) and (258) to display the interlayer
synchronized state XM = XS .
T
The synchronization error is then defined as limT →∞ (1/T ) 0 ∥δ X(t )∥dt.
The linear stability of this solution can be assessed rigorously via the analysis of the linearized system for small δ X, which
reads:
N

δ ẋi = [Df(xM
i ) − σ2 χi Dh2 (xi )]δ xi − σ1
M
Lij Dh1 (xM
j )δ xj , (260)
j =1
where Df, Dh1 , and Dh2 are the Jacobian functions, and XM = g(t ) is the MN state to be targeted. This equation represents the
MSF for the solution corresponding to the identical synchronization between MN and SN. Each of the linear equations (260),
solved in parallel to the N nonlinear equations for the MN (256), corresponds to a set of m conditional Lyapunov exponents at
each pinning configuration (PC). Therefore, each particular PC for which the largest of all such exponents is negative makes
the interlayer synchronous state XM = XS stable.
In Ref. [414], the validity of such a targeting procedure was demonstrated for layers of Rössler oscillators, configured
with Erdős–Rényi (ER) [417] and Barabási–Albert SF [385] topologies. More precisely, a pair of identical master and slave
layers were considered, described by the following set of equations:
 M
ẋi = −yMi − zi ,
M



ẏi = xi + 0.2yMi ,
 M M

żi = 0.2 + zi xi − 7.0 + σ1 zjM − ziM ,
M M
  M  


j∈Ni
ẋi = −ySi − ziS ,

 S

 
ẏi = xSi + 0.2ySi + σ2 (yM
k − yk ),
 S
 S
t k∈T
 
żi = 0.2 + zi xi − 7.0 + σ1 zjS − ziS ,
S S

  S  

j∈Ni
where Ni denotes the neighborhood of the node i, and Tt the set of pinning nodes for which a unidirectional connection is
formed from the master to the slave layers.
The equations above correspond, in the notations of Eqs. (256)–(258), to xM ≡ (xM , yM , z M ), xS ≡ (xS , yS , z S ), h1 (x) =
(0, 0, z ), h2 (x) = (0, y, 0) and f = [−y − z , x + 0.2y, 0.2 + z (x − 7.0)].
By fixing σ1 = σ2 = 1, the Authors of Ref. [414] considered both sequences of pinning nodes following directly the
degree ranking, or following a random ranking (in which the order of appearance of each node is fully random), in order to
compare between both ends of the whole spectrum of possible targeting strategies.
In Fig. 50, the results are shown for the first 200 targeting steps, sufficient in all cases to attain the desired goal dynamics
g(t ). It is easy to see that the maximum Lyapunov exponent λmax , obtained from Eq. (260) and reported in the upper panel
of the figure, always crosses the zero line exactly in correspondence of a vanishing synchronization error, reported in the
lower panel of the figure.
Furthermore, a comparative inspection of Fig. 50 allows to draw two important conclusions: (i) for both SF and ER
topologies, the degree ranking leads to a much better targeting, as compared to the random scheme, and (ii) the relative
Fig. 50. (Color online) Targetability of different topologies. Targeting scheme for ER and SF networks of N = 500 Rössler oscillators. The top panel shows
λmax as a function of the targeting step, for degree ranking and random ranking. The zero level is highlighted as a black horizontal line. The bottom panel
shows the synchronization error vs. the targeting step. The legend contains the symbol codes for the specific topologies considered, as well as for the
specific ranking used for the targeting. All results correspond to ensemble averages over 40 different graph realizations, and, for the random ranking case
over 40 different ranking realizations.
improvement in SF networks is much more evident than in the ER case, indicating that the dynamics of heterogeneous
structures, like the ones encountered in real-world networks, are much easier to be manipulated and targeted.
7. Applications
If the success of a scientific theory can be assessed by its contributions to the understanding of real-world problems,
complex networks can possibly be seen as one of the most prosperous developments in the last decades, having found
applications in a plethora of different contexts [418]. Nevertheless, it is usually recognized that the traditional network
representation is an oversimplification of the corresponding system; for instance, social interactions seldom develop on a
single channel, and more than one relationship can bind pairs of people. It is thus not surprising that multilayer networks
have raised a great expectation, as they should be able to better represent the topology and dynamics of real-world systems.
In the following, we present a review of works dealing with the analysis of real multilayer networks. Most of them were
introduced well before the recent development of the mathematical framework reported here, when Erving Goffman first
proposed his frame analysis social theory in 1974. Due to this, the reader should expect a large heterogeneity in concepts
and solutions.
This section is organized into three main topics: social, technological and economical networks — a synopsis of the main
references is provided in Table 6.
Social networks have captured most of the attention, thanks to the data availability maintained by online communities,
and the inherent multilayer structure of social relationships. Yet, relevant problems can also be found in the two latter
topics, as for instance the analysis of transportation systems and of trade networks. The end of this section is devoted to
other applications, which have hitherto received less attention, but whose relevance, and the relevance of the corresponding
multilayer representations, is expected to grow in the near future: this includes biomedicine, e.g. the analysis of the brain
dynamics taking into account the layers of neural organization, climate (global networks created by correlation between
climatological variables), ecology and psychology.
7.1. Social
7.1.1. Multilayer social networks: an old concept

While the mathematical formulation of multilayer networks is very young, the idea of studying the way social interactions
develop on multiple layers dates back to the seventies. The first of such approaches was probably introduced by Erving
Goffman in 1974, along with the theory of frame analysis [463]. According to this research method, any communication
between individuals (or organizations) is constructed (or framed) in order to maximize the probability of being interpreted
in a particular manner by the receiver. Such framing may differ according to the type of relations between the involved
individuals, several of them potentially overlapping in a single communication: thus, the output of multiple framings can
be interpreted as a multilayer structure.
Table 6
Resume of the main application topics, and related references.
Resume of topics and references
Field Topic References
Pardus: [63,419–422]
Netflix: [423,424]
Flickr: [66,88,425]
Online communities Facebook: [68,426–428]
Youtube: [429]
Other online communities: [54,89,430]
Merging multiple communities: [122,123,
Social
431,432]
Internet [109,110,433]
Citation networks DBLP: [31,33,434–439]
Scottish Community Alliance: [440]
Politics: [68,441]
Other social networks Terrorism: [23]
Bible: [442]
Mobile communication: [443]
Power grids: [25,81,444]
Interdependent systems
Space networks: [445]
Multimodal: [149,184]
Technical
Transportation systems Cargo ships: [446]
Air transport: [16,78]
Other technical networks Warfare: [447]
International Trade Network: [70,71,448]
Trade networks
Maritime flows: [449]
Economy
Interbank market [450]
Organizational networks [451–453]
Biomedicine [454–459]
Climate [24,460]
Other applications
Ecology [64,461]
Psychology [462]
After this first attempt, White et al. proposed an extension of the classical work by Nadel and Fortes [464] to perform
a quantitative analysis of multilayer social networks [465]. A simple three-layer structure is considered, where nodes
represent biomedical researchers specialized in the neural control of food and water intake; the three layers respectively
represent the existence of a bidirectional personal tie, an unidirectional awareness, and the absence of awareness [466].
Notice that this three-layer structure contains information that, in modern complex network theory, would be synthesized
into a single directed network, as the third layer is the complementary of the second. Using this network as a test bed, they
proposed a method for detecting blocks that are coherent across the different layers: in other words, they had just introduced
the first community detection algorithm. Along this line of research, Greiger and Pattison in Ref. [467] analyzed the
multilayer social structure reported in Ref. [468], characterized by business and marriage relationships between important
families from Florence in the fifteenth century. Authors showed how a personal hierarchy, comprising information from both
layers, can be used to create meaningful partitions of the graph.
Finally, it is worth concluding this short review on the classical social applications by presenting the work proposed in
Ref. [469]. Authors argued that an analysis of the health condition in old people, e.g. their resilience to external shocks, can
be performed only when multiple networks are simultaneously considered, like living arrangements (i.e. ties created by the
residence type), family, or organizations like church or voluntary associations.
7.1.2. Online communities

Information and communications technology (ICT) is behind the boost experienced in the last decade by the online
communities. ICT favors users’ interactions through the sharing of multiple contents, from tastes to photos and videos,
and at the same time it fulfills one of the most basic social needs: interpersonal communication. On the other hand, all those
communication exchanges are stored in a digital support: researchers can thus find plenty of information for analyzing
social behaviors, and for representing the wide variety of human interactions under a multilayer framework.
The Pardus social game. One of the most interesting online community data sets has been provided by Pardus, an online game
in which more than 300.000 players live in a virtual, futuristic universe, exploring and interacting with others, pursuing their
personal goals by means of different communication channels ranging from instantaneous one-to-one communications,
e.g. e-mail and personal messages, while others imply a more stable relationship (friendship, cooperation, attacks, and so
forth) [470].
Several papers have used this data set to explore whether particular network topological features were present. For
instance, Ref. [63] extracted a six-layers structure, respectively representing friendship or enmity relations between
Fig. 51. Link overlap (Jaccard coefficient), degree correlation and rank correlation for all pairs of layers in the Pardus social game network. E, enmity; F,
friendship; A, attack; T, trade; C, communication; and B, bounty. Pairs of equal connotation (positive–positive and negative–negative) are marked with a
gray background.
Source: Reprinted figure from Ref. [63]. Courtesy of S. Thurner.
pairs of players, private messages exchange, trading activities, the presence of one-to-one aggressive acts, and placing
head money (bounties) on other players. Different characteristics were found for each layer: specifically, negatively
connoted (i.e. aggressive) actions are characterized by power-law topologies, while positive ones exhibit a higher clustering
coefficient. Furthermore, different layers interact in a non-trivial way: for instance, there is a strong overlap between
communication and friendship networks, but also between communication and aggression, and between communication
and friendship — see Fig. 51. Ref. [419] expands further those results, by considering the time evolution of such a multilayer
network. The network dynamics follows some well-known processes, like triadic closure and network densification.
Furthermore, it seems that the Dunbar conjecture is confirmed [471], as out-degrees are limited at around 150 connections.
More recently, the effect of the user’s gender was tackled in Ref. [420], demonstrating that the structural properties of the
layers are different for men and women, especially when the interaction time scale is considered.
The problem of the triadic closure has more recently been addressed in Ref. [421], through the analysis of a Pardus data
set composed of friendship, communication and trading relationships. Specifically, a network growth model is presented,
in which links are either added to form a closed triangle with probability r, or following a standard preferential attachment
mechanism [385] with probability 1 − r. Such a simple model is yet able to reproduce most of the topological features of
the original multilayer network, thus suggesting that both dynamics compete in the real system.
Finally, Ref. [422] tackles the problem of detecting elite structures in the Pardus network, i.e. subgroups of individuals
that have the ability of influencing the behavior of others. This is accomplished by means of a modified k-core algorithm.
Results indicate that different layer projections can be used to predict how players will perform socially, e.g. their leadership
or wealth.
Netflix. Netflix, Inc. is a US-based company specialized in video rental and online streaming. When users navigate in the
website, for instance to select the next content they want to watch, an advance recommender system [472] analyzes their
tastes and suggests a set of videos expected to be to their liking. In 2006, with the aim of improving this recommendation
system and position it above competitor ones, Netflix disclosed a large data set, including millions of movie ratings assigned
by users [473].
Two works [423,424] have examined this data set by considering its structure as a multilayer network. Specifically, the
initial bipartite network (users can only be connected to movies) is projected into a movie’s network, i.e. where nodes
represent movies, pairwise connected when a single user rated both of them. Furthermore, information about ratings
has been used to organize links in three different layers: positive, negative and mixed ratings. The resulting three layers
have different characteristics, which are lost when the global projection is considered. For instance, positive and negative
networks display a high clustering coefficient, indicating that users tend to like (or dislike) groups of similar movies; on the
other hand, mixed ratings create more heterogeneous structures, characterized by larger but sparser communities.
Flickr. Flickr is a worldly recognized online community, specialized in the hosting and sharing of multimedia content like
photos and videos. In 2013, the popularity of this web service was such that counted more than 87 million registered
users, and 3.5 million new images uploaded daily. Due to the large quantity of information available about interactions
between users, it is not a surprise that several studies have leveraged on Flickr to validate network analyzes, both applied
(e.g. automatic assignation of tags to photos by means of community detection algorithms [474]) and theoretical (as
validating network growing mechanisms against real data [475]).
Recently, three works have focused on the multilayer structure of the Flickr community. While nodes always represent
users, these may be linked through different types of relations. Specifically, Authors identify 9 of them, spanning from direct
user friendship, to common comments or common tag usage.
First of all, two of them [66,425] have introduced the concept of multilayer networks in the design of improved
recommendation systems, i.e. systems able to recommend new friends or relevant contents to target users. Results indicate
that users are more likely to share photos and tags when they are already friends, or in other words, that the social layer
drives the semantic one. Also, there is a strong bidirectionality in social relations: if a user comments on the photo of
another user, the latter reciprocates by commenting a photo of the former person. It is worth noticing that these results, and
especially the relationship between social and semantic connections, could not be inferred if all layers were collapsed into a
single-layer network. Furthermore, time is here an essential ingredient: data is then represented as multilayer time-varying
graphs.
Finally, Ref. [88] focused on the problem of identifying communities of users across the structure. This is accomplished
by firstly generalizing the concept of local clustering coefficient to multilayer networks (see Section 2.2.2) averaging the
metric as measured in each layer. Secondly, the communities are detected through a growing mechanism, in which nodes
are assigned to independent communities when their respective clustering coefficient is higher than a given threshold.
Facebook. Facebook.com is a social website where people can interact and share tastes and comments about their daily
activities. What is uncommon about it, is its popularity: in 2013 it was recognized as the second most visited website in the
world, with 1.230 million registered users, requiring a farm of 50 thousand servers for ensuring its correct operation. As for
other online social networks, many works can be found in the Literature which analyze the topology emerging from user
interactions, as for instance Refs. [476–479]; and possibly Facebook has attracted even more scientific attention due to the
large business possibilities provided by the system.
The analysis of Facebook as a multilayer network has mainly been possible by the work presented in Ref. [426]. In that
reference, Authors have crossed online interaction information with real-world data for a group of 1.640 students from
a private college in the northeast of the United States. The resulting data set comprises five different layers: Facebook
friendship, picture friendship (i.e. two users sharing photos), and roommate, dormmate and groupmate relationships.
Furthermore, the data set is longitudinal, as it includes information at different time moments between 2006 and 2009;
finally, it includes demographic and cultural details of each user.
Taking advantage of the usefulness of such a valuable data set, several works have leveraged on it to perform different
types of analyzes, some of them focusing on its multilayer dimension. Among others, they include the study of different
types of multilayer structures [68,427] up to dynamical processes, like peer influence [428].
Youtube. Youtube.com is a website whose purpose is video sharing, i.e. it allows users to upload personal videos and share
them with the community. In February 2014, it was recognized as the third website in the world in terms of traffic generated.
While few works have focused on the network dimension of the website, Ref. [429] analyzed in a multilayer context
the interactions among Youtube registered users. The resulting network is composed of five layers, each one defining
different relations between pairs of users: friendship, co-contacts (i.e. when two users both add a third one as a friend),
co-subscription (when they follow the same content), co-subscribed (when two users are followed by the same people), and
when they share favorite videos. The tackled problem is the detection of communities of users, sharing the same tastes and
interests. The proposed solution is the extraction of structural features from each layer of the network, and their integration
via Principal Component Analysis [480].
Other online communities. Beside these popular websites, additional (and usually less well-known) online communities
have been studied from the multilayer perspective, representing more specific types of interactions. Here below, some of
those analyzes are reported.
In the first one, Ref. [430] analyzes a Swiss online social network connecting musicians through friendship, such
that subscribers can be classified either as users (followers and fans) or as artists. Three different layers are considered,
respectively created by friendship, communication, and music download. Notably, the three networks share a similar
structure, with a clear presence of homophily and reciprocity.
Ref. [89] studies the social network behind the web portal extradom.pl, specialized in gathering people engaged in
building their own houses in Poland. Eleven different layers are reconstructed from user interactions, including forum groups
and topic creation, or message posting. The interest resides in the analysis of the neighborhood of each user, specifically
through a multilayered neighborhood. Results indicate that, while people are exposed to many different types of relationships,
they tend to focus themselves in just one or two types.
Finally, Ref. [54] tackles the problem of defining node centrality (see Section 2.2.1) by expanding the classical PageRank
algorithm [84], and by using as a test bed the social network of the online community at the University of California, Irvine.
Such network comprises two layers: the first one representing direct instant messaging, the second the bipartite topology
created by users and discussion groups. As it may be expected, taking the multiplex nature of the network into account,
i.e. using a multiplex PageRank, helps uncovering the emergence of rankings of nodes that differ from the rankings obtained
from the projected topology.
Merging multiple social networks. After realizing the wealth of social networks populating Internet, and how information
can straightforwardly be extracted and processed, the reader may wonder about the possibility of merging online networks,
where nodes would represent individuals, and links representing friendship (or other relations) between pairs of them
sitting at different layers, each layer codifying a different social site. While this is in principle feasible, the main problem
resides in the identification of common users across websites, as real names are usually concealed, and the same user can
appear under different pseudonyms, or nicknames.
One way of tackling this problem is by using information extracted from Friendfeed, a social media aggregator. Besides
performing regular activities like posting messages, it also allows users to register their accounts in other systems, thus
allowing the retrieval of the multiple accounts associated to the same user in several social services. Ref. [123] make use of
this website to obtain information from Friendfeed, Twitter and YouTube, creating a network of 7.628 nodes to test different
multilayer network generation algorithms. The same data set is also the source in Refs. [122,431] for node centrality studies.
An alternative approach has been proposed by Ref. [432], where two content-driven websites (Pinterest and Last.fm) are
scanned for connections toward a generic social network, Facebook. The work aimed at confirming social bootstrapping as
a common technique used by novel websites, that is, the act of recommending new social ties depending on the existence
of that link in a different network.
7.1.3. Internet
In the previous part of this section, the World Wide Web has been considered as an instrument to put people in contact,
and therefore as an instrument to drive online social networks. Yet, Internet is also the result of social forces: for instance, the
structure created by hyperlinks is partly the result of connections between real people. Here we present networks created
by hyperlinks between web contents as a social aspect of Internet.
Three papers addressed this problem from a multilayer point of view. Firstly, Refs. [109,110] consider a data set created
by the automatic crawling of 4.986 unique URLs, connected between them by 122.196 hyperlinks. Each web page is also
associated to one or more anchor text terms, i.e. the visible words of an hyperlink. After an initial data filtering, the final
network was composed of 787 pages and 533 terms, each one of them defining a different layer. Upon this network, an
algorithm called TOPHITS was developed, aimed at identifying clusters of similar content web pages and at ranking web
sites by taking into account the multilayer structure of links.
Ref. [433] describes the creation of a university URL-citations network, a network generated by checking the number of
hyperlinks connecting web pages of couples of universities. While this network has a single-layer nature, complementary
information has been added, which creates additional graphs supporting (and partly explaining) the initial structure. For
instance, two universities can be connected when they share the same primary language, when they are located in the same
country, or when they are located within a given physical distance. Results indicate that the topological structure is mainly
explained by language, USA and UK universities forming a main cluster, while the geographical distance is not relevant.
7.1.4. Citation networks

Since the beginning of the complex network theory, citation networks captured a lot of attention, especially as far as
the topology underlying citations between scientific papers was concerned. In this latter case, analysis is motivated, firstly,
by the relative simplicity in obtaining the network structure, and secondly, by the interest in creating new metrics better
assessing the importance of manuscripts and journals.
When moving to a multilayer setting, the main source of information comes from the DBLP (formerly known as DataBase
systems and Logic Programming, and later as Digital Bibliography & Library Project), a computer science bibliography website
hosted by Universität Trier, Germany, nowadays listing more than 2.3 million references.
Several works have been based on such data set, starting from the seminal paper by Cai et al. about community
detection [434]. In this network representation, nodes correspond to Authors, pairwise connected when they have published
a paper in the same conference; the more than 1.000 conferences recognized by DBLP constitute the multilayer structure of
the network. Later works have used the same network to study specific characteristics of the system, like for instance node
centrality assessment and ranking [33,435,436], the discovery of shortest-paths [437], and the identification of coherent
subgraphs [438].
On the other hand, the DBLP data set has also been used as a test ground for more general multilayer network analysis. For
instance, Ref. [439] proposes the use of higher order tensors algebra to multilayer and time evolving networks, i.e. an offline
tensor analysis (OTA) as an extension of the celebrated Principal Component Analysis [481,482]. Also, Ref. [31] proposes the
extension of several well-known metrics, e.g. the degree of a node, as well as new ones, as the case of the dimension relevance
of a node in a layer.
7.1.5. Other social interactions

In addition to online human interactions and citations, other types of multilayer social interactions can be found in the
Literature. Here, we briefly report them, as an example of the variety of goals and approaches that can be accomplished by
means of the multilayer network paradigm.
As a first example, Ref. [440] describes the Scottish Community Alliance, a network composed of local communities of
people, deriving from the ancient Highlanders communities. Such a system has the structure of a network of networks, as
communities are organized in larger groups at a regional level; but it is also organized in different layers, according to
their main topics: ‘‘Community Energy Scotland’’, ‘‘Community Woodlands Association’’, the ‘‘Federation of City Farms and
Community Gardens’’ and the ‘‘Scottish League of Credit Unions’’.
Politics is the focus of Ref. [68] and specifically the network created by call voting: nodes thus represent US senators,
and pairs of them are connected if they shared similar voting patterns across one year. The analysis of multiple years is
thus represented as a multilayer structure. Authors demonstrate the usefulness of such a representation by performing
a community analysis, and showing how the community structure can be used to infer information about the global
political situation. Specifically, 9 communities are identified, which are associated to important historical situations as the
Compromise of year 1850, the beginning of the American Civil War, or the Great Depression.
Another example can be found in Ref. [23] which analyzes the Noordin Top Terrorist Network, composed of 78 known
terrorists active in Indonesia [483]. Each network is defined by the interactions between pairs of individuals, i.e. trust,
operational, communication and business relations.
The social network built by Saint Paul along its biblical voyage through the Aegean sea is characterized in Ref. [442],
which considers the two layers corresponding to strong and weak ties. While the former (including close friends and
family) was usually considered to be essential for spreading risky ideas (thus including recruitment to, and spreading
of the new religion), results suggest that Saint Paul mainly achieved his mission through a large network of weak social
links.
Ref. [443] is concerned about the process of social communication using different technologies as phone calls, or short
text messages (SMS). The considered data set has been created from billing information corresponding to a large European
metropolitan area, from March 26 to May 31, 2012. It contains more than 63 millions phone-call records, and 20 millions
SMS records. The two layers do not perfectly overlap, with a 30% (8%) of users using only phone calls (SMS). Moreover, most
of the links, i.e. communications between two people, exclusively occur by means of phone calls, while the share of people
exchanging only SMS is as low as 21%.
Finally, Ref. [441] analyzed the networking between representatives from 168 leading groups on national health policy
located in Washington DC during year 2003. Three overlapping networks, i.e. layers, are considered: communication,
coalition overlap and issue overlap.
7.2. Technical systems
7.2.1. Interdependent systems

One of the most promising fields of application of the multilayer framework is the study of technical interdependent
systems, i.e. those sets of systems in which the correct functioning of one of them strongly depends on the status of the
others. In other words, the failure of one element does not only provoke the breakdown of similar elements, but also
affects the dynamics of other coupled systems. This leads to an iterative process of cascading failures, as summarized before
in Section 4, that has a devastating effect on the network stability. In a natural way, different types of systems can be
represented by layers, connected between them by dependency links [25,177].
Electrical power grids have been one of the first targets of such analysis, due to their high interconnectedness with both
networks of different nature, like communication networks providing synchronization between different stations [25], and
of similar kind, like neighboring power grids [81]. The devastating effects of such cascading failure dynamics has been firstly
studied in Ref. [25]. Specifically, Authors analyzed the electrical blackout that affected the major part of Italy on 28 September
2003. The shutdown of power stations directly led to the failure of nodes in the Internet communication network, which in
turn caused further power station breakdowns.
Two years later, Brummitt et al. in Ref. [81] analyzed the interactions between different power grids, specifically between
the three main regions constituting the US network — Western, Eastern, and Texas. Adding some connectivity between the
different regions is beneficial, as it suppresses the largest cascade in each system by providing alternative paths. Yet, higher
levels of interconnectivity have a negative impact, both because they open pathways for neighboring networks to inflict large
cascades, and because new interconnections increase capacity and total possible load, which fuels even larger cascades. In
the same spirit, Ref. [444] again considered the interaction between the power grid and the supporting communication
network by focusing on the problem of detecting the most central, i.e. the most vulnerable, node. Authors focused on the
Maricopa County, the most densely populated county of Arizona, USA. The resulting network comprised 70 power plants,
470 transmission lines, and 42.723 fiber links.
Beyond power grids, it is worth noticing the completely different problem tackled in Ref. [445]. There, Authors introduced
space-based networks, i.e. a technology that allows the sharing of spacecraft on-orbit resources, such as data storage,
processing, and downlink [484] — see Fig. 52. Each spacecraft in the network can have different subsystem composition and
functionality, and can rely on others to complement (or back-up) its equipment. Thus, in this case, each layer represents a
spacecraft, with links connecting different layers representing the possibility of sharing capabilities. Various design insights
are derived and discussed, and the capability to perform trade-space analysis with the proposed approach for various
network characteristics is indicated.
7.2.2. Transportation systems

Transportation systems are natural candidates for a multilayer network representation. Firstly because connections
between different nodes, e.g. direct flights or trains, are explicit, that is, their existence is published, requiring little effort for
the network construction. Secondly, because transportation systems are usually interconnected, with a multilayer approach
Fig. 52. Example of a space-based network. The two light gray boxes represent two spacecrafts, connected through a high-capacity communication system.
Each one of them is composed of different modules, e.g. the Telemetry, Tracking and Command (TTC): when one of them fails, the corresponding spacecraft
can rely on the other to backup that function, and to fulfill its mission.
thus being the natural choice for intermodality studies. This latter aspect, i.e. the multimodal nature of transportation, has
been tackled in two recent papers. In the first one, a two-layer structure has been created by merging the world wide port
and airport networks [184], with the aim of studying the global robustness against random attacks. While previous studies
have considered the resilience of coupled systems when such a coupling was of a random nature [25,35], there Authors
focused on the level of similarity between layers, given for instance the fact that well-connected airports are more connected
to central ports. Such a similarity is assessed by means of two metrics, namely, the inter degree–degree correlation and
the interclustering coefficient. Simulations executed using the topology of such a transportation system reveal that, as the
networks become more intersimilar, the system becomes significantly more robust to random failure, as one network is
able to back-up the other, when needed.
In the second publication [149], a multilayer structure is used to model the Indian air and train transportation networks,
the nodes of the latter being stations, with pairs of them being connected if within one stop distance. Pairs of nodes belonging
to different layers, i.e. airport–station pairs, are connected whenever a direct road access between both of them is present.
It is worth noticing that both networks have a very different nature. While the airport network is small (with 78 nodes)
and presents a scale-free disassortative structure, the rail one is much larger (6.769 stations) and has a topology strongly
constrained by the geography, implying a narrow degree distribution and an assortative topology. Yet, such heterogeneity
enables a good performance of the global network in terms of navigability.
As for the analysis of individual transportation networks, a few examples can be found in the Literature, in which a
multilayer structure is used to model different aspects of the transportation process. For instance, Ref. [446] analyzed the
network created by worldwide cargo ship movements; while the 90% of world trade is carried by sea, it is noteworthy that
little attention has been devoted to this transportation mode from the field of complex networks. The multilayer structure is
given by the three most common classes of cargo ships, i.e. container ships, bulk dry carriers and oil tankers, which usually
connect different ports and travel following different patterns. For instance, bulk dry carriers and oil tankers tend to move in
a less regular manner between ports than container ships. The understanding of the specific patterns of movement created
by each type of ship, instead of analyzing the topological projection of this multilayer network, is motivated by the problem
of invasive species spreading: bulk dry carriers and oil tankers often sail empty, thus exchanging large quantities of ballast
water and favoring the spreading process.
Finally, the air transport network [19] has also been studied alone, but taking into account the multilayer structure
created by different airlines. As demonstrated in Ref. [78], the global network is the result of the interactions between two
main types of layers: those created by major airlines on the one side, and those generated by low cost (or low fares) carriers
on the other. The former are characterized by a hub-and-spoke topology, with one central hub receiving and dispatching
all flights; the latter by a point-to-point configuration, of a more random nature. The topological properties of the global
network, as for instance its low diameter or the high clustering coefficient, only appear when several layers of different
nature are merged together. While an analysis of the topology of connections may be relevant, at least from a theoretical
point of view, the study of the dynamics taking place on top of it is of high interest for the air transport community, as
it would allow tackling problems like the propagation of delays through the system, or the effects of such dynamics on
the movement of passengers [19]. Ref. [16] tackles the problem of passenger rescheduling when a fraction of the links,
i.e. of flights, is randomly removed. A simple dynamical model is executed using two networks: the single-layer projected
network, and the whole multilayer structure. In the latter, airlines incur in a cost whenever a passenger needs to change
layer, thus effectively reducing the flexibility of the system. Results indicate that the layered structure is associated to a
lower resilience of the system against such kind of perturbations. In other words, the use of a projection is a simplification
that results in an over-estimation of the resilience of the air transport network, as it does not take into account the cost of
moving between layers — see Fig. 53 for more details. At the intersection between the topological and dynamical analysis of
the air transport network, Ref. [218] focused on the communicability, i.e. the number of possible routes that two nodes can
choose to communicate with each other [216]. By aggregating the communicability corresponding to each departure node,
i.e. considering the broadcasting of information to passengers, and by varying the degree of interactions between different
layers, it is possible to create airport rankings that takes into account the multilayer structure or, in other words, quantifies
the effects of the airline structure on the movement of passengers. The most central node thus varies: while the airport of
Paris CdG is the most central in Europe when layers are disconnected (main consequence of the extension of the Air France
network), this role is passed to London Stansted and Frankfurt when increasing the coupling.
n=0 n=1 n=2

<fnf > <fnf > <fnf >
0.3 1 0.3 1 0.3 1
0.9 0.9 0.9
0.25 0.8 0.25 0.8 0.25 0.8
0.7 0.7 0.7
0.2 0.6 0.2 0.6 0.2 0.6
ftol
ftol
ftol
0.5 0.5 0.5
0.15 0.4 0.15 0.4 0.15 0.4
0.3 0.3 0.3
0.1 0.2 0.1 0.2 0.1 0.2
0.1 0.1 0.1
0.05 0 0.05 0 0.05 0
10-3 10-2 10-1 1 10-3 10-2 10-1 1 10-3 10-2 10-1 1
p p p
<fol > <fol > <fol >
0.3 1 0.3 1 0.3 1
0.9 0.9 0.9
0.25 0.8 0.25 0.8 0.25 0.8
0.7 0.7 0.7
0.2 0.6 0.2 0.6 0.2 0.6
ftol
ftol
ftol
0.5 0.5 0.5
0.15 0.4 0.15 0.4 0.15 0.4
0.3 0.3 0.3
0.1 0.2 0.1 0.2 0.1 0.2
0.1 0.1 0.1
0.05 -3 0 0.05 -3 0 0.05 -3 0
10 10-2 10-1 1 10 10-2 10-1 1 10 10-2 10-1 1
p p p
<fsl > <fsl > <fsl >
0.3 1 0.3 1 0.3 1
0.9 0.9 0.9
0.25 0.8 0.25 0.8 0.25 0.8
0.7 0.7 0.7
0.2 0.6 0.2 0.6 0.2 0.6
ftol
ftol
ftol
0.5 0.5 0.5

0.15 0.4 0.15 0.4 0.15 0.4
0.3 0.3 0.3
0.1 0.2 0.1 0.2 0.1 0.2
0.1 0.1 0.1
0.05 -3 0 0.05 -3 0 0.05 -3 0
10 10-2 10-1 1 10 10-2 10-1 1 10 10-2 10-1 1
p p p
Fig. 53. (Color online) Outcome of the re-scheduling process in the multilayer air transport network, as a function of the probability of link failure, p, and
load tolerance, ftol . Each column displays (from top to bottom): the average fraction of passengers that cannot fly, that of those that are re-scheduled in
other layers, and that for those re-scheduled within the original layer. Each column accounts for the possibility of scheduling passengers on paths with
length 0 (left), 1 (center), and 2 (right).
© 2013, With kind permission from Springer Science and Business Media.
7.2.3. Other technical systems

As a final note on technological systems, it is worth noticing Ref. [447], which analyzes the problem of establishing a
reliable communication network between heterogeneous elements, and specifically between different warfighters in an
army. Thanks to the technological progress, each soldier and vehicle in the battlefield can be connected to others by means
of different communication channels, each one having specific characteristics: being wireless or not, bandwidth, range,
etc. When these different systems are considered as heterogeneous layers, the battlefield can be represented by a single
multilayer network — see Fig. 54. The challenge is then to ensure a high end-to-end Quality of Service (QoS), i.e. that two
units can share information when needed, making use of different communication channels.
7.3. Economy
7.3.1. Trade networks

The International Trade Network (ITN), also known as the World Trade Web (WTW), is defined as the graph of all
import/export relationships between world countries. In recent years, it has been extensively analyzed from the complex
network theory point of view, due to its importance: for instance, its topology can be used to assess the likelihood that
economic shocks might be transmitted between countries [485].
Traditional network studies have considered the ITN at the aggregate level, in which links represent the presence of a
trade relation, irrespectively of the commodity actually traded. Nevertheless, the introduction of the concept of multilayer
network has enabled the unfolding of the aggregate ITN such that each layer represents import and export relationships for
a given commodity class, according to some standard classification scheme.
This latter approach has been conducted in Refs. [70,71], by using a network composed of 162 countries (nodes) and 97
layers, evolving yearly from 1992 to 2003. In the former work, Authors considered different topological metrics and how
they differ across the layers: among others, node centralities (and thus, country rankings) and connected components. On
the other hand, the latter work focused on the identification of communities, highlighting the fact that commodity-specific
communities have a degree of granularity which is lost in the aggregate analysis. A similar study was performed in Ref. [448],
Fig. 54. (Color online) Example of an heterogeneous mobile communication network. The battlefield can be seen as a set of resources, e.g. tanks, soldiers,
or even satellites, which require to continuously transmit tactical information to the others. Due to the different characteristics and capabilities of each
resource, multiple communication networks are deployed (as, for instance, satellite, UHF or VHF links), making the system a multilayer network.
Source: Reprinted figure from Ref. [447]. Courtesy of P. Lubkowski.
but relying on a different data set: the UNCOMTRADE, collected by the United Nations, which includes information for 94
countries starting from year 1962.
A different approach to trade analysis has been proposed in Ref. [449] regarding the study of maritime flows. Extending
the large research body devoted to the study of maritime flows by means of standard complex networks [446,486,487],
Ref. [449] presents a model where ports (nodes) are connected through five different layers, representing the five main
categories of commodities exchanged in maritime business: liquid bulk (i.e. crude oil, oil products, chemicals, etc.), solid
bulk (like aggregates, cement, or ores), containers, passengers/vehicles and general cargo. Data have been extracted from
the Lloyd’s Voyage Records, covering the months of October and November 2004. An inspection of this disaggregated network
suggests that most networks measures, like centrality or clustering, exhibit strong sensitivity to the type of commodity
traded in a given set of ports. The most diversified ports are on average, the bigger and more dominant they are in the
network, while they also connect over greater physical distances than more specialized ports — see Fig. 55.
7.3.2. Interbank market

The 2008 financial crisis has raised the interest of the scientific community toward the understanding of the complex
system emerging from the interaction between financial institutions and markets [488–490]. Most contributions have
focused on the interbank market, where nodes represent financial institutions, and links credit relations between two
counterparties. Networks are usually directed and weighted, such that directionality identifies the borrower and the lender,
and the weight of the link represents the loan amount.
Ref. [450] is hitherto the only work analyzing this network from a multilayer point of view. Authors utilize a unique
database of supervisory reports of Italian banks to the Banca d’Italia, including all bilateral exposures of all Italian banks,
broken down by maturity and by the secured and unsecured nature of the contract. Each layer in the network represents
a different type of exposure, namely: unsecured overnight, unsecured short-term (less than 12 months), unsecured long-
term (more than 12 months), secured short-term and secured long-term. The analysis of these five layers shows a high
heterogeneity between them, with the topology of the overall network largely reflecting the one of the unsecured overnight
layer. This layer is especially important from the policy-making point of view, as it is the focus of monetary policy operations
in several jurisdictions. Notably, such layer is characterized by a high persistence of links over time, such that, if two banks
are connected at a given time, they will have a high probability of being connected in the future. This may have important
consequences in the stability of the system, in terms of perturbations contagion.
7.3.3. Organizational networks

A third active field of research in economy is the study of organizational networks, i.e. the study of how different
organizations interact, by exchanging information, money, or other resources, with the aim of collectively produce a good or
a service. While traditionally such interactions had a hierarchical structure, in recent years more complex topologies have
emerged, in which the power of the decision making process is spread among many participants [491]. Once again, complex
networks stand for as the most suitable instrument for the analysis of the resulting interaction structures [492,493]. While
Fig. 55. (Color online) Commodity diversity and nodal regions in the maritime flow network. Nodes represent ports, pairwise connected by links when a
commodity is transported between them. Different commodities are considered, thus making this graph a multilayer network. In order to represent this
heterogeneity, links are colored according to the number of commodity types traveling between ports. Furthermore, both nodes and links sizes represent
the corresponding traffic share.
© 2013, with permission from Elsevier.
the importance of the multilayer nature of these networks was recognized almost three decades ago [494,495], only recently
this problem has been mathematically tackled.
One of the first works analyzing organizational networks as multilayer structures is Ref. [451]. There, Authors analyzed
more than 6.000 Research & Development and more than 6.500 Management & Development alliances involving 1.000 biotech
firms in the United States. The time window considered covers over 30 years, in which data were available quarterly, thus
enabling the construction of a 120-layers temporal network. The underlying hypothesis is that the actions of a given firm
can be better understood if one takes into account changes in the network structure, like the firm position (centrality). The
main result suggests that biotech firms act following a process of preferential attachment, i.e. organizations are more likely
to form ties with organizations of similar institutional and structural status.
The multiplex paradigm has also been used to study the interactions between organizations using Information and Com-
munications Technology (ICT) firms for development goals, i.e. the use of ICTs as a tool for empowering less developed
countries [452]. Specifically, two nodes can be connected when the corresponding organizations collaborate in implemen-
tation or knowledge-sharing projects, thus creating a two-layers network. In Ref. [452], the studied data set comprised 323
projects between years 1997 and 2005, 211 of which were implementation projects and 112 knowledge-sharing projects,
for a total of 752 organizations. Results indicate that the existence of a link in one layer is influenced by several factors: the
presence of the corresponding link in the other layer, the presence of a common third-party, or the centrality of connected
nodes, among others.
Within organizational networks analysis, of special interest has been Corporate Governance, defined as the study of the
interrelations between ownership and control of a firm. Stockholders invest their money in a company whose strategic
decisions are taken by the top management: as both of them are embedded in a network (for instance, a single investor
may hold shares of multiple companies), the resulting structure can be seen as a multilayer network composed of two
layers, respectively ownership and control. Thus, a multilayer analysis may give a unique perspective on the relevance of
management rather than ownership in the company’s policy. Such an approach has been addressed in Ref. [453], focusing
on the shareholding and board structures in the year 2013 for 273 quoted companies on the Italian Stock Market. The main
claim is that corporate governance cannot be understood unless ownership and management are taken simultaneously,
as the two layers are characterized by a small overlap: while the core of shareholding is composed of banks and financial
institutions, the most central nodes in the management layer are large industrial companies.
7.4. Other applications
7.4.1. Biomedicine
The introduction of the complex network methodology in biomedicine caused a kind of a small revolution, as it made
possible a real systems medicine: not only studying the elements constituting a living being, but analyzing how they
communicate and interact in a systemic way. Two are the main fields where the network approach has been successful.
On the one side, the characterization of how the main constituents of the cell are organized: specifically, genes [496,497],
proteins [380,498] and metabolites [499]. On the other side, and from a higher level perspective, one may find all the works
in neuroscience that deals with the complexity of the brain as a network, both anatomically [500] and functionally [501].
Up to now, these networks, e.g. the ones created by interactions between genes, proteins and metabolites, have been
considered in an independent fashion. Nevertheless, they are clearly not independent: genes control the production of
proteins, which in turn catalyze metabolic reactions. Malfunctions in cells are seldom due to just one element, being
instead the result of multiple interactions at different levels. It has been recently suggested that a multilayer approach, i.e. a
representation of the cell composed of different layers for each one of these elements, may yield relevant knowledge about
the normal dynamics, as well as about the appearance of pathological conditions [454]. The final aim is called personalized
medicine, i.e. a medicine focused on the individual and proactive in nature, which would allow a significant improvement in
health care by customizing the treatment according to each patient needs [455]. In spite of its importance, such multilayer
approach has not yet been achieved.
While a full multilayer representation of the cell has not yet been constructed, there are several situations in which
different networks can be combined or analyzed together. For instance, molecule interaction networks can be extracted from
different organisms, the objective being the identification of modules that have been preserved across different evolution
steps. Following this line of research, in Ref. [456] Authors found common modules by means of a spectral clustering
algorithm extending the Perron–Frobenius theorem. Among the resulting shared modules, some examples include the Mini
Chromosome Maintenance complex, which plays an important role in eukaryotes DNA replication [502], or the V-ATPase, an
ancient and highly preserved enzyme that performs different functions in eukaryotes cells [503]. It has also been proposed
the multilayer analysis of the protein–protein interaction network by considering the type of relation connecting pairs of
them: physical when they can bind together to form a larger complex, and genetic if one of the two proteins regulates the
other, i.e. by triggering the activation (or repression) of the gene responsible for the production of the target protein [457].
Another example of a multilayer approach emerges when a set of data, representing the same molecules, are extracted
from different conditions, or are analyzed by different means. These two approaches are explored in Refs. [458,459].
Specifically, Ref. [458] works with a multilayer structure composed of 130 co-expression networks, in which each layer
represents a different experimental condition. The aim was the identification of recurrent heavy subgraphs, i.e. subsets
of densely interconnected nodes that consistently appear in the different layers. While this is similar to a community
identification task in a single layer, the simultaneous analysis of multiple layers is expected to enhance signal–noise
separation. In Ref. [459], an initial data set of gene expressions is used to create two different networks: firstly, a standard co-
expression network, where pairs of genes are connected whether a relationship (e.g. statistical correlation) is found between
their expression levels; and secondly, an exon co-splicing network, in which nodes represent exons and the edge weights
represent correlations between the inclusion rates of two exons across all samples in the data set. These two networks are
not independent, but indeed form a two-layers system, as each gene contains several exons. Furthermore, transcription and
splicing factors are expected to be consistent among related genes, thus giving rise to coupled modules. This last hypothesis
has been confirmed by comparing the obtained modules with data sets of gene functions.
7.4.2. Climate
The last decade has witnessed an increasing interest toward the application of complex network theory to climate
science. Networks are usually reconstructed such that nodes correspond to spatial grid points of the Earth, and links are
added between pairs of nodes depending on the degree of statistical interdependence between the corresponding pairs of
anomaly time series taken from the climate data set. The graph approach then allows a topological and dynamical analysis
of the climate system over many scales, i.e. from local properties of a single point and its influence on the whole planet, to
global network measures associated to stabilization mechanisms [504–506].
Ref. [24] constructs a multilayer network, where each layer represents the Earth’s atmospheric condition at a given
geopotential height; nodes are associated to pairs of coordinates (i.e. latitude and longitude), and pairs of them are connected
with a weight defined by the zero-lag linear correlation between the time series corresponding to the two nodes. In this way,
the multilayer network provides a representation of the geostrophic wind field as well as large-scale convection processes
in the troposphere and the lower stratosphere. By analyzing its topology, and especially the centrality of nodes (by means of
adapted degree, closeness and betweenness centralities), it is possible to extract relevant knowledge about the atmosphere
dynamics, without considering fields of temperature, moisture content or other relevant climatological variables that are
not straightforward to obtain.
A different approach is proposed in Ref. [460], where a two-layer network is constructed to analyze sea–air interactions.
Nodes in the two layers respectively codify positions (pairs of latitudes and longitudes), and links are once more weighted
according to a zero-lag linear correlation between time series associated to each node: geopotential heights for points in the
air, and surface sea temperature for points in the sea. By analyzing this network, it is possible to recover the coupled Asian
monsoon and Walker circulations, known as the Indian and Pacific Ocean (GIP) model.
7.4.3. Ecology
In this subsection, we concentrate on two paperworks that analyzed, from a multilayer point of view, the interactions
between animals and the structure of an aquifer system, respectively.
In the first one, the interaction network of a group of 12 female baboons in the De Hoop Nature Reserve, South Africa,
is considered [64]. Specifically, three different types of interactions were recorded along 10 years, by means of human
observations: agonistic, spatial and grooming interactions. This remarkable data set has then been used to build a multilayer
structure, in which each layer represents each one of those social aspects. Beyond the complete, or control, network, two
more have been reconstructed, by monitoring the social changes resulting from the death of two subjects, specifically a
dominant and a low-ranked female. This allowed Authors to compare the topological properties of the three networks,
specifically their clustering coefficients, and analyze how social ties are modified by the two events. The loss of the dominant
female resulted in an instability in the dominance hierarchy, which was compensated for by adjustments in the spatial
association network. On the other hand, the loss of the low-ranked female had little effect on the social network. This
suggests that dominance serves to regulate the interactions of all group members.
A completely different application is proposed in Ref. [461], where the aquifer system of the Orageval catchment in
France is studied. In order to correctly describe the structure of this system, a two-layer network representation has been
reconstructed, where the two layers respectively map the Oligocene and the Eocene geological layers. After geophysical
and drilling investigations, aimed at assessing the general structure of the aquifer system, electrical resistivity tomography
was used to reconstruct the connectivity patterns at different spatial resolutions. When combined with temperature
and hydraulic head information, this multilayer structure provides the researcher with valuable information about the
stream–aquifer interface.
7.4.4. Psychology
While psychology has not traditionally contemplated complex networks as a tool for analyzing human behavior, it is
worth noticing the study proposed in Ref. [462]. In this work, the symbolism of dreams and their interpretation, i.e. oneirol-
ogy [507], is studied by means of networks: nodes represent dream symbols, pairwise connected according to the overlap
in their meaning (i.e. the similarity in their corresponding interpretation documents). A different network is created for
each considered culture (English, Chinese and Arabic), for then constructing a network of networks by connecting common
symbols. Findings suggest that sentiments connected with a given interpretation can be associated with communities in the
network, and that communities can be themselves clustered into three main categories of positive, negative, and neutral
dream symbols — thus revealing a complex multiscale structure — see Fig. 56.
8. Conclusions and open questions
After having revisited the main theory and applications of multilayer networks, we feel our duty to draw a few concluding
remarks, and pinpoint questions that remain still open to future progresses, at the same time addressing the reader to some
issues and problems that we consider of key relevance, and that we believe it is more than likely that would attract soon
the attention of scientists in the area.
In Section 2 we started with trying to establish a common notation (and some basic definitions) for multilayer networks,
this way joining forces that have been spent in the last years to describe the structure of several complex systems. The main
inspiration was trying to prove that the different terminologies reported so far could be, in fact, encompassed within a single
framework. However, there is still a lot of work ahead in order to set a proper and comprehensive mathematical formalism,
and possibly a long way to go. Among the different open problems to be solved, we highlight the following three: (i) the need
of setting up other metric concepts that could possibly affect relevant parameters of the systems, such as the betweenness,
the vulnerability and the efficiency amongst others; (ii) the need of gathering a better knowledge and understanding on the
mathematical relationships among the values of centrality parameters when measured on each single network’s layer and
Fig. 56. (Color online) Communities in the dream interpretation network: English (left), Chinese (center), and Arabic (right). Symbols with the highest
eigenvector centrality are used to label each community. The complete networks are illustrated in the upper right corners of each figure, with colors
corresponding to the communities (colors are not intended to match across different networks).
when, instead, measured on the whole multilayer structure; (iii) the need of paying effort for extending other quantifications
pertinent to the dynamics of single-layer networks, such as, for instance, those regarding controllability and observability.
In Section 3, we followed up by reviewing the different models that have been proposed for generating multilayer
networks endowed with the topological structures observed in real world systems. Possibly, this is the part of state of the
art that still lies in its infancy. Definitely, and much more than being confident of, this issue is of the utmost importance.
Retrieving the mechanisms at the basis of what is ubiquitously seen in social, technological and biological environments
will certainly contribute to uncover the intimate physical processes that ultimately determine the way networked systems
shape their interactions. Multilayer and multiplex networks encode indeed, significant more information in their structure
with respect to their single-layers taken in isolation. A direct way to extract information from these structures is to look at
correlations. For multiplex networks, these can be represented by degree correlations between different layers, or overlap
of the links, or correlations in the activities of the nodes. For networks of networks, the degrees of freedom are even more
pronounced since the interlinks between the nodes in different layers can be assigned using a large variety of rules. In our
Section 3, we have tried to give an overview of generative models using both a growing multilayer network framework,
and the framework of an ensemble of networks generating unbiased null models. Growing multiplex network models can
be used to generate multiplexes with positive correlation in the degrees of the nodes, when the preferential attachment is
linear, or tunable (positive or negative) degree correlation when the preferential attachment is non linear. On the other side,
multiplex network ensembles using multilinks can be used to generate multiplex networks with tunable levels of overlap
of the links incident to each node, that cannot be generated using the configuration model in each layer of the multiplex.
At variance, in Section 4 we attempted to review what are probably the most studied problems in multilayer networks,
i.e. the issues connected to the overall network’s resilience (or fragility) to external perturbations, and to percolation. Such
processes, indeed, once defined on multilayer networks, display many relevant and unexpected properties that cannot be
simply reported to the case of a single layer. As soon as interdependencies between the nodes in different layers are taken
into account, one of the most important questions is to characterize their robustness properties, i.e. their response to random
damage. In systems like complex infrastructures, we observe that an increased level of interdependencies can be responsible
for an increased fragility of the system that might be affected by cascades of failure events. On the other hand, in a network
with interdependencies between the nodes of different layers, the percolation transition can be discontinuous, indicating a
discontinuous emergence/disruption of the mutually connected component. At the same time, the response to damage of
these networks generates eventually a disrupting avalanche of failures events. The investigation of the effects of multiplexity
into the robustness of networks has revealed that when the networks are uncorrelated and each node has at least one
interdependency link, the transition is always expected to be discontinuous and characterized by avalanches of failures.
The nature (and structural characteristic) of the interlink might also affect the robustness properties of the network. For
instance, in the context of networks of networks, if any node is linked to its replica node and there is a fixed supernetwork,
the size of the mutually connected component is independent of the structure of the supernetwork as long as it is connected,
while in the case of nodes randomly linked to other nodes in other layers, the structure of the supernetwork matters.
Moreover, if the nodes are only connected to their replica ones, either all the layers percolate, or none does. In the case
in which the nodes in one layer are linked to random nodes in other layers, one has that layers with higher superdegrees are
more fragile than layers with lower superdegrees. Moreover, in the presence of partial interdependence or overlap of the
links, the transition can become continuous, and the response to damage less disruptive. This opens the challenging problem
of characterizing the general structural conditions that can enhance resilience and robustness of the entire system. While
for infrastructure and man made networks it is plausible to assume that the design of the system has not taken into account
the role of interdependencies and has therefore resulted in fragile structures, in biological system such as the cell (or the
brain) we are more inclined to believe that evolution must have driven the system toward a relatively robust configuration,
at least finding a tradeoff between robustness and adaptability of the entire system.
In Section 5 we have shown that spreading processes in multilayer networks constitute possibly one of the best suited
dynamical frameworks to unveil the effects that the coupling between networks has on their functioning. For instance,
linear diffusion being the most simple dynamical setup allows to monitor the transition between two limiting regimes (one
in which the networks behave independently and another corresponding to a coordinated functioning) with the transition
between these two phases appearing to be abrupt according to the spectrum of the Laplacian. The use of more refined
diffusion processes such as random walkers allows to go one step further and device tools for ranking the importance of the
nodes, and monitoring it as a function of the coupling between networks. The study of even more realistic diffusion processes,
such as those mimicking the transport of information packets, on interconnected networks has been also addressed,
although more results on this line are certainly to be expected. In this regard, the transition to congestion in information
networks is of key importance.
In addition to the report on diffusion processes, Section 5 paid attention to the problem of disease spreading on multilayer
networks, where several related disease dynamics propagate on separated network layers, or both social information and
infection information simultaneously diffuse and affect each other. Compared with the case of single-layer networks, this
new framework can greatly change the threshold of disease outbreak, which is usually shown in the form of a phase
transition between the healthy and epidemic phases. More research regarding this important issue is needed since SIR
and SIS models seem to behave differently. In particular, SIR dynamics allows system states in which the epidemic regime
only occurs in one of the networks (while not affecting the others) whereas for the SIS dynamics the propagation in one
network automatically induces the epidemic in all the remaining networks. Besides the prediction of the epidemic threshold,
we also emphasized studies that take advantage of the multilayer benchmark to couple the spreading dynamics with the
development of voluntary prevention measures to control and eradicate the infection risk. Along this research line, there is
still room for many proposals considering interconnected networks. As it is known, social dynamics, including information
propagation and opinion formation, plays a non-trivial role in the risk perception under an epidemic threat. Apart from
studying the emergence of individual contention measures under disease outbreaks, the same approach seems to be
adequate to explore dynamical prevention measures in the Internet, so to control, in a dynamical way, the attack of viruses.
As we have seen in Section 5, evolutionary games are another fascinating field of research that can benefit from the
multilayer framework. At variance with games in single-layer networks, evolutionary games on different network layers
can be coupled via either the utility function or the strategy of the nodes. We summarized the main influence of these
kinds of inter-layer coupling on the maintenance of collective cooperation behavior both in pair-wise games (such as the
Prisoner’s Dilemma) and n-person ones (such as the Public Goods Game). The reviewed studies clearly point out that the
different coupling between networks enriches the enhancement of cooperative behaviors provided by spatial reciprocity.
In spite of the great achievements, there are still unexplored problems related to evolutionary games that merit further
attention. The most obvious issue is extending to multilayer networks other types of games like the Ultimatum game [508,
509], the Rock–Scissor–Paper game [510], the Naming game [511,512] and the Collective-risk social dilemma [513]. These
games usually focus on the emergence of other collective states such as, fairness, species diversity, cyclical dominance,
language evolution and prevention of dangerous climate change. Within the novel multilayer framework, one can get a
better understanding on the impact that the interaction among different networks plays on the emergence of collective
states. For instance, during the campaign of mitigating climate change, the mutual impact of different local systems is crucial.
If the total contribution of these systems is less than the expected target, it may cause a cascade of reducing investment
leading to the tragedy of the commons [514]. Experimental research is another unexplored field of evolutionary games on
multilayer networks. Previous experiments on single-layer networks have unveiled that heterogeneous topology does not
promote cooperation [515], which is inconsistent with theoretical predictions [341]. Naturally, if the multilayer interaction
is involved, how cooperation trait varies will become of particular importance.
In Section 6 we have given a rather complete overview of the current results about synchronization in multilayer
networked systems. Starting from the simplest case of only two layers, we have first reported on the enhancement of
synchronization when the topological evolution of a network happens through a sequence of commutative layers. We have
then demonstrated how one can in fact obtain a fully general treatment, without the need to impose any constraint on the
layers and their mutual relations. Other specific cases we considered are coexisting layers, hypergraphs, interconnected
networks, multiplex networks and bipartite graphs. Finally, we discussed recent results on the problem of targeting
dynamical trajectories of a system. The new approach we described for general network evolution opens several possibilities
for research areas in the near future. In particular, it should be noted that by the application of Eqs. (252) and (255) one can
effectively specify the evolution of the Laplacian eigenvalues and eigenvectors during the switch from one layer to another.
Thus, one can impose any kind of transient behavior in the network. This method can be exploited to achieve finer control
on time-evolving complex networks, with potential applications to many areas of social impact.
Further, the role of symmetry on the dynamical properties of multilayer networks is currently completely unexplored.
Studies exist describing the importance of topological symmetries for synchronization, but they all deal with single-layer
graphs. The addition of new layers could lead to substantially different results, which need to be investigated. Notice that
the two cases of temporal layers and spatial layers are likely to involve different mathematics, with potentially different
conclusions.
Another very relevant subject of investigation that will certainly attract attention is the study of how a multilayer
distribution of links can spontaneously emerge as a consequence of adaptation of the interactions between constituents of a
system. In Refs. [516–518] it was indeed shown that, when considering a single-layer network of phase oscillators and when
the interaction weights are themselves dynamical variables evolving by means of differential equations explicitly coupled
with the state of the network’s nodes, the asymptotic emergence of a synchronous state is associated with a structural self-
organization into a modular and scale-free topology. The extension of this formalism to the case of multilayer interactions
between oscillators is in progress, and promises to unveil relevant information on how a multilayer structure of connections
emerges spontaneously in connection with the setting of specific collective dynamics.
Finally, in Section 7 we have reviewed the current literature on applications of the multilayer formalism and represen-
tation to several cases of relevance in social sciences, technological systems and economy. We firmly believe that in the
future biology and biomedicine will greatly take advantage of a multilayer representation, although these two fields of re-
search have until now lag behind. This is mainly due to the difficulty associated with the recording of real data in living
organisms, especially with the high resolution needed to discriminate the physical layers. This is, for instance, the case of
the human brain: while the human cortex is composed of six layers, each one processing information in a distinctive way,
actual recording technologies (e.g. EEG or MEG) can only detect the activity of the top-most layers. When functional rep-
resentations (i.e. representing the activity of the brain in terms of information processing) will be merged with structural
ones (e.g. the connectome, that is, the structure of fibers connecting different parts of the brain), the result will be a complex
multilayer network, describing both how and why information moves. Beyond the brain, it is commonly accepted that cells
are composed of different networks interacting between them: from the RNA, mRNA, up to proteins and metabolites. While
these networks have been deeply studied in separated and isolated fashion, many pathological conditions generate from
the wrong interaction of the different networks (layers). Adopting the multilayer paradigm will then provide an elegant
mathematical way of representing these cross-interactions, paving the way for a deeper comprehension of the structure
and function of cells.
Note added in proof
After editing the present Report, we learned about the publication of other recent and important contributions which
relate to subjects and arguments treated in our Sections. In the following, we provide a list of these additional References
that the reader may find useful to consult, together with the number of the Section they refer to.
Section 2:
M. De Domenico, and V. Nicosia, ArXiv e-prints (2014) 1405.0425v1; I. Laut, C. Räth, L. Wörner, V. Nosenko, S.K. Zhdanov,
J. Schablinski, D. Block, H.M. Thomas, G.E. Morfill, Phys. Rev. E 89 (2014) 023104; R.J. Sánchez-García, E. Cozzo, Y. Moreno,
Phys. Rev. E 89 (2014) 052815; M. De Domenico, M.A. Porter, A. Arenas, ArXiv e-prints (2014) 1405.0843; C.W. Loe,
H.J. Jensen, ArXiv e-prints (2014) 1406.2205; A. Solé-Ribalta, M. De Domenico, S. Gómez, A. Arenas, WebSci ’14 Proceedings
of the 2014 ACM conference on Web science (2014) 149–155.
Section 4:
F. Tan, Y. Xia, ArXiv e-prints (2014) 1405.4342; Z. Su, L. Li, H. Peng, J. Kurths, J. Xiao, Y. Yang, Sci. Rep. 4 (2014) 5413;
L. Daqing, J. Yinan, K. Rui, S. Havlin, Sci. Rep. 4 (2014) 5381; B. Kotnis, ArXiv e-prints (2014) 1406.2106.
Section 5:
E. Valdano, L. Ferreri, C. Poletto, V. Colizza, ArXiv e-print (2014) 1406.4815; F. Bagnoli, E. Massaro, ArXiv e-prints (2014)
1405.5348; W. Wang, M. Tang, H. Yang, Y. Do, Y.C. Lai, G.W. Lee, Sci. Rep. 4 (2014) 57; B. Wang, Z. Pei, L. Wang, ArXiv e-
prints (2014) 1405.1573; B. Sun, B. Leonard, P. Ronhovde, Z. Nussinov, ArXiv e-prints (2014) 1406.7282; J. Sanz, Ch.-Y. Xia,
S. Meloni, Y. Moreno, ArXiv e-prints (2014) 1402.4523; M. Salehi, R. Sharma, M. Marzolla, D. Montesi, P. Sivari, M. Magnani,
ArXiv e-prints (2014) 1405.4329; C. Granell, S. Gomez, A. Arenas, ArXiv e-prints (2014) 1405.4480; A. Chmiel, P. Klimek,
S. Thurner, ArXiv e-prints (2014) 1405.3801.
Section 6:
L.M. Pecora, F. Sorrentino, A.M. Hagerstrom, T.E. Murphy, R. Roy, Nat. Comm. 5 (2014) 4079; V. Nicosia, P.S. Skardal,
V. Latora, A. Arenas, ArXiv e-prints (2014) 1405.5855v1; M. Asllani, D.M. Busiello, T. Carletti, D. Fanelli, G. Planchon, ArXiv
e-prints (2014) 1406.6401.
Section 7:
D. Hristova, M. Musolesi, C. Mascolo, Proceedings of the 8th AAAI International Conference on Weblogs and Social Media
(ICWSM’14), Ann Arbor, Michigan, USA.
Furthermore, a part of the material quoted in our presentation was taken from the ArXiv public repository, and we warn
the reader that it is very likely that those Manuscripts could have appeared meanwhile on peer-reviewed Journals. For
instance, we ourselves were made aware of publication of some of those sources, that we gladly report here below:
Our Ref. [98] was published as Proc. Nat. Acad. Sci. USA 111 (2014) 8351.
Our Ref. [154] was published as Phys. Rev. E 89 (2014) 062814.
Our Ref. [226] was published as Phys. Rev. E 89 (6) (2014) 062813.
Our Ref. [257] was published as PLoS ONE 9 (3) (2014) e92200.
Our Ref. [323] was published as Phys. Rev. E 89 (6) (2014) 062818.
Acknowledgments
We would like to acknowledge gratefully all colleagues with whom we maintained (and are currently maintaining)
interactions and discussions on the topic covered in our report.
In particular, we would like to thank J.A. Almendral, A. Amann, R. Amritkar, K. Anand, R.F.S. Andrade, F.T. Arecchi,
A. Arenas, S. Assenza, V. Avalos-Gaytán, A. Ayali, N. Azimi-Tafreshi, A. Baronchelli, M. Barthélemy, K.E. Bassler, E. Ben Jacob,
P. Bonifazi, J. Bragard, J.M. Buldú, J. Burguete, M. Campillo, A. Cardillo, T. Carroll, D. Cellai, M. Chavez, M. Courbage, A. Díaz-
Guilera, L. Dall’Asta, D. De Martino, A.L. Do, S.N. Dorogovtsev, P. Echenique, E. Estrada, L.M. Floría, J. Flores, S. Fortunato,
J. Gao, E. García, S. Gómez, A. García del Amo, J.P. Gleeson, C. Gracia-Lazaro, C. Grebogi, P. Grigolini, B. Guerra, R. Gutiérrez,
A. Halu, M. Hassler, S. Havlin, H.G.E. Hentschel, A. Hramov, D.-U. Hwang, Q. Jin, H. Kim, A. Koronovski, J. Kurths, V. Latora,
S. Lepri, I. Leyva, F. Lillo, R. Livi, Y. Liu, E. Lopez, P. Lubkowski, S. Luccioli, H. Mancini, M. Marsili, N. Masuda, D. Maza, S. Meloni,
J.F.F. Mendes, R. Meucci, B. Min, G. Mindlin, R.J. Mondragon, Y. Moreno, O. Moskalenko, S. Mukherjee, F. Naranjo, A. Navas,
M. Newmann, V. Nicosia, J. García-Ojalvo, C.J. Pérez-Vicente, P. Panzarasa, D. Papo, L.M. Pecora, M. Perc, P. Pin, A. Pikovski,
A.N. Pisarchick, A. Politi, J. Poncela, F. del Pozo, I. Procaccia, A.A. Rad, I. Reinares, D. Remondini, C. Reyes-Suárez, M. Rosenblum,
R. Roy, S. Ruffo, G. Russo, F.D. Sahneh, A. Sánchez, D. de Santos-Sierra, S.E. Schaeffer, C. Scoglio, J.R. Sevilla Escoboza,
K. Showalter, L. Solá, S. Solomon, H.E. Stanley, R. Stoop, S. Strogatz, K. Syamal Dana, E. Strano, A. Szolnoki, S. Thurner,
A. Torcini, Z. Toroczkai, V. Tsioutsiou, D. Vilone, L. Wang, C-y. Xia, Y. Zhang, K. Zhao, and J. Zhou.
The uncountable number of stimulating and inspired discussions with them (and their sharing with us of unpublished
results on the subject) are largely responsible for having spurred and encouraged us to provide the present account on this
rapidly growing field of research.
At the same time, we also would like to gratefully acknowledge all the discussions we had (in meetings, Conferences,
Workshops, and personal visits) with other colleagues who are not mentioned in the above list (and, for that, we really
apologize), which equally inspired our efforts, and opened up our minds in a way that contributed, eventually and
substantially, to the realization of the present survey.
This work was partly supported by the Spanish MINECO under projects FIS2011-25167, FIS2012-38949-C03-01, and
FIS2012-38266-C02-01; by the European FET project MULTIPLEX ‘‘Foundational research on multilevel complex networks
and systems’’ (Reference Number 317532); by the Comunidad de Aragon (FENOL group); the Brazilian CNPq through the
PVE project of the Ciencia Sem Fronteiras. C.D.G. acknowledges support by EINS, Network of Excellence in Internet Science,
via the European Commission’s FP7 under Communications Networks, Content and Technologies, grant No. 288021; J.G.G is
supported by MINECO through the Ramon y Cajal program. Z.W. acknowledges the National Natural Science Foundation of
China (Grant No. 11005047).
References
[1] S.H. Strogatz, Exploring complex networks, Nature 410 (6825) (2001) 268–276. http://dx.doi.org/10.1038/35065725.
[2] R. Albert, A.-L. Barabási, Statistical mechanics of complex networks, Rev. Mod. Phys. 74 (1) (2002) 47–97.
http://dx.doi.org/10.1103/RevModPhys.74.47.
[3] S.N. Dorogovtsev, J.F. Mendes, Evolution of networks, Adv. Phys. 51 (4) (2002) 1079–1187. http://dx.doi.org/10.1080/00018730110112519.
[4] M. Newman, The structure and function of complex networks, SIAM Rev. 45 (2) (2003) 167–256. http://dx.doi.org/10.1137/S003614450342480.
[5] D. Watts, Small Worlds: The Dynamics of Networks Between Order and Randomness, Princeton University Press, Princeton, NJ, USA, 1999.
[6] S. Bornholdt, H.G. Schuster (Eds.), Handbook of Graphs and Networks: From the Genome to the Internet, John Wiley & Sons, Inc, New York, NY, USA,
2003.
[7] S.N. Dorogovtsev, J.F. Mendes, Evolution of Networks: From Biological Nets to the Internet and WWW, Oxford University Press, Oxford, 2003.
[8] R. Pastor-Satorras, A. Vespignani, Evolution and Structure of the Internet: A Statistical Physics Approach, Cambridge University Press, Cambridge,
2004.
[9] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, D.-U. Hwang, Complex networks: structure and dynamics, Phys. Rep. 424 (4–5) (2006) 175–308.
http://dx.doi.org/10.1016/j.physrep.2005.10.009.
[10] G. Szabó, G. Fath, Evolutionary games on graphs, Phys. Rep. 446 (4) (2007) 97–216. http://dx.doi.org/10.1016/j.physrep.2007.04.004.
[11] A. Arenas, A. Díaz-Guilera, J. Kurths, Y. Moreno, C. Zhou, Synchronization in complex networks, Phys. Rep. 469 (3) (2008) 93–153.
http://dx.doi.org/10.1016/j.physrep.2008.09.002.
[12] S. Fortunato, Community detection in graphs, Phys. Rep. 486 (2010) 75–174. http://dx.doi.org/10.1016/j.physrep.2009.11.002.
[13] M. Barthélemy, Spatial networks, Phys. Rep. 499 (1–3) (2011) 1–101. http://dx.doi.org/10.1016/j.physrep.2010.11.002.
[14] S. Wasserman, K. Faust, Social Networks Analysis, Cambridge University Press, Cambridge, 1994.
[15] J. Scott, Social Network Analysis: A Handbook, SAGE Publications, London, 2000.
[16] A. Cardillo, M. Zanin, J. Gómez-Gardeñes, M. Romance, A. García del Amo, S. Boccaletti, Modeling the multi-layer nature of the european air transport
network: Resilience and passengers re-scheduling under random failures, Eur. Phys. J. Spec. Top. 215 (1) (2013) 23–33.
http://dx.doi.org/10.1140/epjst/e2013-01712-8.
[17] R. Criado, M. Romance, M. Vela-Pérez, Hyperstructures, a new approach to complex systems, IJBC 20 (03) (2010) 877–883.
http://dx.doi.org/10.1142/S0218127410026162.
[18] R. Criado, B. Hernández-Bermejo, M. Romance, Efficiency, vulnerability and cost: an overview with applications to subway networks worldwide,
IJBC 17 (07) (2007) 2289–2301. http://dx.doi.org/10.1142/S0218127407018397.
[19] M. Zanin, F. Lillo, Modelling the air transport with complex networks: a short review, Eur. Phys. J. Spec. Top. 215 (2013) 5–21.
http://dx.doi.org/10.1140/epjst/e2013-01711-9.
[20] M. Berlingerio, M. Coscia, F. Giannotti, A. Monreale, D. Pedreschi, Foundations of multidimensional network analysis, in: 2011. International
Conference on Advances in Social Networks Analysis and Mining, 2011, pp. 485–489. http://dx.doi.org/10.1109/ASONAM.2011.103.
[21] M. De Domenico, A. Solé-Ribalta, E. Cozzo, M. Kivelä, Y. Moreno, M.A. Porter, S. Gómez, A. Arenas, Mathematical formulation of multilayer networks,
Phys. Rev. X 3 (2013) 041022. http://dx.doi.org/10.1103/PhysRevX.3.041022.
[22] M. Kivelä, A. Arenas, M. Barthélemy, J.P. Gleeson, Y. Moreno, M.A. Porter, Multilayer networks, ArXiv e-prints. arXiv:1309.7233.
[23] F. Battiston, V. Nicosia, V. Latora, Structural measures for multiplex networks, Phys. Rev. E 89 (2014) 032804.
http://dx.doi.org/10.1103/PhysRevE.89.032804.
[24] J. Donges, H. Schultz, N. Marwan, Y. Zou, J. Kurths, Investigating the topology of interacting networks, Eur. Phys. J. B 84 (4) (2011) 635–651.
http://dx.doi.org/10.1140/epjb/e2011-10795-8.
[25] S.V. Buldyrev, R. Parshani, G. Paul, H.E. Stanley, S. Havlin, Catastrophic cascade of failures in interdependent networks, Nature 464 (7291) (2010)
1025–1028.
[26] J. Gao, S.V. Buldyrev, H.E. Stanley, S. Havlin, Networks formed from interdependent networks, Nat. Phys. 8 (1) (2012) 40–48.
[27] R. Axelrod, The dissemination of culture: a model with local convergence and global polarization, J. Confl. Resolut. 41 (2) (1997) 203–226.
[28] L. Solá, M. Romance, R. Criado, J. Flores, A. García del Amo, S. Boccaletti, Eigenvector centrality of nodes in multiplex networks, Chaos 23 (3) (2013)
033131.
[29] J.P. Holme, Temporal networks, PR 519 (3) (2012) 97–125.
[30] M. Coscia, Multidimensional network analysis (Ph.D. thesis), Universitá Degli Studi Di Pisa, Dipartimento di Informatica, 2012.
[31] M. Berlingerio, M. Coscia, F. Giannotti, A. Monreale, D. Pedreschi, Multidimensional networks: foundations of structural analysis, World Wide Web
16 (5–6) (2013) 567–593.
[32] M. Berlingerio, F. Pinelli, F. Calabrese, Abacus: frequent pattern mining-based community discovery in multidimensional networks, Data Min. Knowl.
Discov. 27 (3) (2013) 294–320.
[33] M. Coscia, G. Rossetti, D. Pennacchioli, D. Ceccarelli, F. Giannotti, ‘‘you know because i know’’: a multidimensional network approach to human
resources problem, in: Proc. of the 2013 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining, ASONAM ’13,
ACM, New York, NY, USA, 2013, pp. 434–441.
[34] M. Kurant, P. Thiran, Layered complex networks, Phys. Rev. Lett. 96 (2006) 138701.
[35] R. Parshani, S.V. Buldyrev, S. Havlin, Interdependent networks: reducing the coupling strength leads to a change from a first to second order
percolation transition, Phys. Rev. Lett. 105 (4) (2010) 048701. http://dx.doi.org/10.1103/PhysRevLett.105.048701.
[36] R. Criado, J. Flores, A. García del Amo, J. Gómez-Gardeñes, M. Romance, A mathematical model for networks with structures in the mesoscale, Int. J.
Computer Math. 89 (3) (2012) 291–309.
[37] C. Berge, Hypergraphs: Combinatorics of Finite Sets, North-Holland, 1989.
[38] D.P. Hypergraphs, in: R. Graham, M. Grötschel, L. Lovász (Eds.), Handbook of Combinatorics, North-Holland, 1995, pp. 381–432.
[39] J.J., Hypernetworks for reconstructing the dynamics of multilevel systems, in: European Conference on Complex Systems, Saïd Business School,
University of Oxford, Oxford, 2006.
[40] E. Estrada, J.A. Rodríguez-Velázquez, Subgraph centrality and clustering in complex hyper-networks, Physica A 364 (2006) 581–594.
[41] E.V. Konstantinova, V.A. Skorobogatov, Application of hypergraph theory in chemistry, Discrete Math. 235 (2001) 365–383.
[42] M. Karonski, T. Luczak, Random hypergraphs, in: D. Miklós, V.T. Sós, T. Szõnyi (Eds.), in: Combinatorics, Paul Erdös is Eighty, vol. 2, Bolyai Society
Mathematical Studies, Budapest, 1996, pp. 283–293.
[43] J.A. Rodríguez, On the laplacian spectrum and walk-regular hypergraphs, Linear Multilinear Algebra 51 (2003) 285–297.
[44] B. Guerra, J. Poncela, J. Gómez-Gardeñes, V. Latora, Y. Moreno, Dynamical organization towards consensus in the Axelrod model on complex networks,
Phys. Rev. E 81 (2010) 056105.
[45] S. Gómez, A. Díaz-Guilera, J. Gómez-Gardeñes, C.J. Pérez-Vicente, Y. Moreno, A. Arenas, Diffusion dynamics on multiplex networks, Phys. Rev. Lett.
110 (2013) 028701.
[46] G. D’Agostino, A. Scala, Networks of Networks: The Last Frontier of Complexity, in: Understanding Complex Systems, Springer, 2014.
[47] E. Cozzo, M. Kivelä, M. De Domenico, A. Solé, A. Arenas, S. Gómez, M.A. Porter, Y. Moreno, Clustering coefficients in multiplex networks, ArXiv e-prints.
arXiv:1307.6780.
[48] B. Min, S.D. Yi, K.-M. Lee, K.-I. Goh, Network robustness of multiplex networks with interlayer degree correlations, Phys. Rev. E 89 (2014) 042811.
[49] G. Bianconi, Statistical mechanics of multiplex networks: entropy and overlap, Phys. Rev. E 87 (6) (2013) 062806.
[50] C.D. Brummitt, K.M. Lee, K.I. Goh, Multiplexity-facilitated cascades in networks, Phys. Rev. E 85 (4 Pt 2) (2012) 045102.
[51] D. Cellai, E. Lopez, J. Zhou, J.P. Gleeson, G. Bianconi, Percolation in multiplex networks with overlap, Phys. Rev. E 88 (5) (2013) 052811.
[52] E. Cozzo, A. Arenas, Y. Moreno, Stability of Boolean multilevel networks, Phys. Rev. E 86 (3 Pt 2) (2012) 036115.
[53] E. Cozzo, R.A. Banos, S. Meloni, Y. Moreno, Contact-based social contagion in multiplex networks, Phys. Rev. E 88 (5) (2013) 050801.
[54] A. Halu, R.J. Mondragon, P. Panzarasa, G. Bianconi, Multiplex pagerank, PLoS ONE 8 (10) (2013) e78293.
[55] J.Y. Kim, K.I. Goh, Coevolution and correlated multiplexity in multiplex networks, Phys. Rev. Lett. 111 (5) (2013) 058702.
[56] K.-M. Lee, J.Y. Kim, W. kuk Cho, K.-I. Goh, I.-M. Kim, Correlated multiplexity and connectivity of multiplex random networks, New J. Phys. 14 (3)
(2012) 033027.
[57] V. Nicosia, G. Bianconi, V. Latora, M. Barthélemy, Growing multiplex networks, Phys. Rev. Lett. 111 (2013) 058701.
[58] V. Nicosia, R. Criado, M. Romance, G. Russo, V. Latora, Controlling centrality in complex networks, Sci. Rep. 2 (2012) 218.
[59] A. Solé-Ribalta, M. De Domenico, N.E. Kouvaris, A. Díaz-Guilera, S. Gómez, A. Arenas, Spectral properties of the laplacian of multiplex networks, Phys.
Rev. E 88 (2013) 032807.
[60] O. Yağan, V. Gligor, Analysis of complex contagions in random multiplex networks, Phys. Rev. E 86 (2012) 036103.
[61] G. Ghoshal, V. Zlatić, G. Caldarelli, M.E. Newman, Random hypergraphs and their applications, Phys. Rev. E 79 (6 Pt 2) (2009) 066118.
[62] V. Zlatić, G. Ghoshal, G. Caldarelli, Hypergraph topological quantities for tagged social networks, Phys. Rev. E 80 (2009) 036118.
[63] M. Szell, R. Lambiotte, S. Thurner, Multirelational organization of large-scale social networks in an online world, Proc. Natl. Acad. Sci. USA 107 (31)
(2010) 13636–13641.
[64] L. Barrett, S.P. Henzi, D. Lusseau, Taking sociality seriously: the structure of multi-dimensional social networks as a source of information for
individuals, Proc. Natl. Acad. Sci. USA 367 (1599) (2012) 2108–2118.
[65] M. Berlingerio, M. Coscia, F. Giannotti, A. Monreale, D. Pedreschi, The pursuit of hubbiness: analysis of hubs in large multidimensional networks,
J. Comput. Sci. 2 (3) (2011) 223–237.
[66] P. Kazienko, K. Musial, T. Kajdanowicz, Multidimensional social network in the social recommender system, IEEE Trans. Syst. Man Cybern. 41 (4)
(2011) 746–759.
[67] P. Kazienko, K. Musial, E. Kukla, T. Kajdanowicz, P. Bródka, Multidimensional social network: model and analysis, in: P. Jedrzejowicz, N.T. Nguyen,
K. Hoang (Eds.), ICCCI 1, in: Lecture Notes in Computer Science, vol. 6922, Springer, 2011, pp. 378–387.
[68] P.J. Mucha, T. Richardson, K. Macon, M.A. Porter, J.P. Onnela, Community structure in time-dependent, multiscale, and multiplex networks, Science
328 (5980) (2010) 876–878.
[69] P. Pattison, S. Wasserman, Logit models and logistic regressions for social networks, ii: multivariate relationships, Br. J. Math. Stat. Psychol. 52 (1999)
169–193.
[70] M. Barigozzi, G. Fagiolo, D. Garlaschelli, Multinetwork of international trade: a commodity-specific analysis, Phys. Rev. E 81 (2010) 046104.
[71] M. Barigozzi, G. Fagiolo, G. Mangioni, Identifying the community structure of the international-trade multi-network, Physica A 390 (11) (2011)
2051–2066.
[72] D.S. Bassett, M.A. Porter, N.F. Wymbs, S.T. Grafton, J.M. Carlson, P.J. Mucha, Robust detection of dynamic community structure in networks, Chaos 23
(1) (2013).
[73] V. Carchiolo, A. Longheu, M. Malgeri, G. Mangioni, Communities unfolding in multislice networks, in: Complex Networks, Springer, Berlin, Heidelberg,
2011, pp. 187–195.
[74] P.J. Mucha, M.A. Porter, Communities in multislice voting networks, Chaos 20 (2010) 041108.
[75] A. Allard, P.-A. Noël, L.J. Dubé, B. Pourbohloul, Heterogeneous bond percolation on multitype networks with an application to epidemic dynamics,
Phys. Rev. E 79 (2009) 036113.
[76] J. Hindes, S. Singh, C.R. Myers, D.J. Schneider, Epidemic fronts in complex networks with metapopulation structure, Phys. Rev. E 88 (2013) 012809.
[77] A. Vazquez, Spreading dynamics on heterogeneous populations: multitype network approach, Phys. Rev. E 74 (2006) 066114.
[78] A. Cardillo, J. Gómez-Gardeñes, M. Zanin, M. Romance, D. Papo, F. del Pozo, S. Boccaletti, Emergence of network features from multiplexity, Sci. Rep.
3 (2013) 1344.
[79] R.J. Sánchez García, E. Cozzo, Y. Moreno, Dimensionality reduction and spectral properties of multilayer networks, ArXiv e-prints. arXiv:1311.1759.
[80] E.A. Leicht, R.M. D’Souza, Percolation on interacting networks, ArXiv e-prints. arXiv:0907.0894.
[81] C.D. Brummitt, R.M. D’Souza, E.A. Leicht, Suppressing cascades of load in interdependent networks, Proc. Natl. Acad. Sci. USA 109 (12) (2012)
E680–689.
[82] P. Bonacich, Power and centrality: a family of measures, Amer. J. Sociol. 92 (5) (1987) 1170–1182.
[83] J. Aguirre, D. Papo, J.M. Buldu, Successful strategies for competing networks, Nat. Phys. 9 (4) (2013) 230–234. http://dx.doi.org/10.1038/nphys2556.
[84] S. Brin, L. Page, The anatomy of a large-scale hypertextual web search engine, in: Computer Networks and ISDN Systems, Elsevier Science Publishers
B. V, 1998, pp. 107–117.
[85] D.J. Watts, S.H. Strogatz, Collective ‘dynamics of ’small-world’ networks, Nature 393 (6684) (1998) 440–442.
[86] R. Luce, A. Perry, A method of matrix analysis of group structure, Psychometrika 14 (2) (1949) 95–116.
[87] V. Latora, M. Marchiori, Economic small-world behavior in weighted networks, Eur. Phys. J. B 52 (2003) 249–263.
[88] P. Bródka, K. Musial, P. Kazienko, A method for group extraction in complex social networks, in: M. Lytras, P. Ordonez De Pablos, A. Ziderman,
A. Roulstone, H. Maurer, J. Imber (Eds.), Knowledge Management, Information Systems, E-Learning, and Sustainability Research, in: Communications
in Computer and Information Science, vol. 111, Springer, Berlin, Heidelberg, 2010, pp. 238–247.
[89] P. Bródka, P. Kazienko, K. Musiał, K. Skibicki, Analysis of neighbourhoods in multi-layered dynamic social networks, Int. J. Computat. Intell. Syst. 5
(3) (2012) 582–596.
[90] L.da F. Costa, F. Rodrigues, G. Travieso, P. Boas, Characterization of complex networks: a survey of measurements, Adv. Phys. 56 (2007) 167–242.
[91] V. Latora, M. Marchiori, Efficient behavior of small-world networks, Phys. Rev. Lett. 87 (19) (2001) 198701.
http://dx.doi.org/10.1103/PhysRevLett.87.198701.
[92] R.G. Morris, M. Barthélemy, Transport on coupled spatial networks, Phys. Rev. Lett. 109 (12) (2012) 128703.
[93] P. Van Mieghem, Graph Spectra for Complex Networks, Cambridge University Press, Cambridge, 2012.
[94] A. Brouwer, W. Haemers, Spectra of Graphs, Springer, Amsterdam, 2012.
[95] W. Haemers, Interlacing eigenvalues and graphs, Linear Algebra Appl. 226 (1995) 593–616.
[96] F. Radicchi, A. Arenas, Abrupt transition in the structural formation of interconnected networks, Nat. Phys. 9 (2013) 717–720.
[97] R. Criado, J. Flores, E. Garcia, A.G. del Amo, M. Romance, L. Sola, A perron-frobenius theory for block matrices and tensor calculus of multiplex
networks, Preprint 1,2014, pp. 1–20.
[98] M. De Domenico, A. Solé-Ribalta, S. Gómez, A. Arenas, Random walks on multiplex networks, ArXiv e-prints. arXiv:1306.0519.
[99] M.A. Porter, J.-P. Onnela, P.J. Mucha, Communities in networks, Notices Amer. Math. Soc. 56 (2009) 1082–1097.
[100] M.E.J. Newman, M. Girvan, Finding and evaluating community structure in networks, Phys. Rev. E 69 (2004) 026113.
[101] R. Milo, S. Shen-Orr, S. Itzkovitz, N. Kashtan, D. Chklovskii, U. Alon, Network motifs: simple building blocks of complex networks, Science 298 (5594)
(2002) 824–827.
[102] S.S. Shen-Orr, R. Milo, S. Mangan, U. Alon, Network motifs in the transcriptional regulation network of escherichia coli, Nat. Genet. 31 (1) (2002)
64–68.
[103] T.G. Kolda, B.W. Bader, Tensor decompositions and applications, SIAM Rev. 51 (3) (2009) 455–500.
[104] D.M. Dunlavy, T.G. Kolda, W.P. Kegelmeyer, Multilinear algebra for analyzing data with multiple linkages, in: J. Kepner, J. Gilbert (Eds.), Graph
Algorithms in the Language of Linear Algebra, Fundamentals of Algorithms, SIAM, Philadelphia, 2011, pp. 85–114.
[105] E. Acar, B. Yener, Unsupervised multiway data analysis: a literature survey, IEEE Trans. Knowl. Data Eng. 21 (1) (2009) 6–20.
[106] C.D. Martin, M.A. Porter, The extraordinary svd, Am. Math. Mon. 119 (10) (2012) 838–851.
[107] J. Carroll, J.-J. Chang, Analysis of individual differences in multidimensional scaling via an n-way generalization of ‘‘Eckart-Young’’ decomposition,
Psychometrika 35 (3) (1970) 283–319.
[108] R. Harshman, Foundations of the parafac procedure: models and conditions for an explanatory multi-modal factor analysis, UCLA Working Papers
in Phonetics 16.
[109] T. Kolda, B. Bader, The TOPHITS model for higher-order web link analysis, in: Proc. of Link Analysis, Counterterrorism and Security 2006, 2006,
pp. 1–12.
[110] T.G. Kolda, B.W. Bader, J.P. Kenny, Higher-order web link analysis using multilinear algebra, in: ICDM 2005: Proc. of the 5th IEEE International
Conference on Data Mining, 2005, pp. 242–249.
[111] T.G. Kolda, J. Sun, Scalable tensor decompositions for multi-aspect data mining, in: Proc. of the 8th IEEE International Conference on Data Mining,
ICDM 2008, 2008, pp. 363–372.
[112] E. Acar, D.M. Dunlavy, T.G. Kolda, Link prediction on evolving data using matrix and tensor factorizations, in: Proc. of the 2009 IEEE International
Conference on Data Mining Workshops, ICDMW ’09, IEEE Computer Society, Washington, DC, USA, 2009, pp. 262–269.
[113] G. Menichetti, D. Remondini, P. Panzarasa, R.J. Mondragón, G. Bianconi, Weighted multiplex networks, ArXiv e-prints. arXiv:1312.6720.
[114] A. Barrat, M. Barthélemy, R.P.-S.A. Vespignani, The architecture of complex weighted networks, PNAS 101 (2004) 3747–3752.
[115] M. Barthélemy, B. Gondran, E. Guichard, Spatial structure of the internet traffic, Physica A 319 (2003) 633–642.
[116] E. Almaas, B. Kovacs, T. Vicsek, Z.N. Oltvai, A.-L. Barabási, Global organization of metabolic fluxes in the bacterium escherichia coli, Nature 427 (2004)
839–843.
[117] V. Nicosia, V. Latora, Measuring and modelling correlations in multiplex networks, ArXiv e-prints. arXiv:1403.1546v1.
[118] S.N. Dorogovtsev, J.F.F. Mendes, Evolution of Networks: From Biological Nets to the Internet and WWW, Oxford University Press, 2003.
[119] C.I.D. Genio, H. Kim, Z. Toroczkai, K.E. Bassler, Efficient and exact sampling of simple graphs with given arbitrary degree sequence, PLoS ONE 5 (2010)
e10012.
[120] H. Kim, C.I.D. Genio, K.E. Bassler, Z. Toroczkai, Constructing and sampling directed graphs with given degree sequences, New J. Phys. 14 (2) (2012)
023012.
[121] V. Nicosia, G. Bianconi, V. Latora, M. Barthélemy, Non-linear growth and condensation in multiplex networks, ArXiv e-prints. arXiv:1312.3683.
[122] M. Magnani, L. Rossi, The ml-model for multi-layer social networks, in: ASONAM, IEEE Computer Society, 2011, pp. 5–12.
[123] M. Magnani, L. Rossi, Formation of multiple networks, in: A. Greenberg, W. Kennedy, N. Bos (Eds.), Social Computing, Behavioral-Cultural Modeling
and Prediction, in: Lecture Notes in Computer Science, vol. 7812, Springer, Berlin, Heidelberg, 2013, pp. 257–264. http://dx.doi.org/10.1007/978-3-
642-37210-0_28.
[124] P. Wang, G. Robins, P. Pattison, E. Lazega, Exponential random graph models for multilevel networks, Social Networks 35 (2013) 96–115.
[125] M. De Domenico, A. Solé-Ribalta, E. Omodei, S. Gómez, A. Arenas, Centrality in interconnected multilayer networks, ArXiv e-prints. arXiv:1311.2906.
[126] S. Funk, V.A.A. Jansen, Interacting epidemics on overlay networks, Phys. Rev. E 81 (2010) 036118.
[127] V. Marceau, P.-A. Noël, L. Hébert-Dufresne, A. Allard, L.J. Dubé, Modeling the dynamical interaction between epidemics on overlay networks, Phys.
Rev. E 84 (2011) 026105.
[128] A. Patrinos, S. Hakimi, Relations between graphs and integer-pair sequences, Discrete Math. 15 (4) (1976) 347–358.
[129] B. Söderberg, General formalism for inhomogeneous random graphs, Phys. Rev. E 66 (2002) 066121.
[130] B. Söderberg, Random graphs with hidden color, Phys. Rev. E 68 (2003) 015102.
[131] B. Söderberg, Properties of random graphs with hidden color, Phys. Rev. E 68 (2003) 026107.
[132] B. Söderberg, Random graph models with hidden color, Acta Phys. Pol. B (2003) 5085.
[133] A. Barbour, G. Reinert, The shortest distance in random multi-type intersection graphs, Random Struct. Algorithm 39 (2) (2011) 179–209.
[134] S. Melnik, M.A. Porter, P.J. Mucha, J.P. Gleeson, Dynamics on modular networks with heterogeneous correlations, Chaos 24 (2) (2014) 023106.
[135] P.W. Holland, K.B. Laskey, S. Leinhardt, Stochastic blockmodels: first steps, Social Networks 5 (2) (1983) 109–137.
[136] S.E. Fienberg, M.M. Meyer, S.S. Wasserman, Statistical analysis of multiple sociometric relations, J. Amer. Statist. Assoc. 80 (389) (1985) 51–67.
[137] Y.J. Wang, G.Y. Wong, Stochastic blockmodels for directed graphs, J. Amer. Statist. Assoc. 82 (397) (1987) 8–19.
[138] M.E.J. Newman, E.A. Leicht, Mixture models and exploratory analysis in networks, Proc. Natl. Acad. Sci. USA 104 (23) (2007) 9564–9569.
[139] G. Robins, P. Pattison, Y. Kalish, D. Lusher, An introduction to exponential random graph (p)models for social networks, Social Networks 29 (2007)
173–191.
[140] G. Robins, T.A.B. Snijders, P. Wang, M. Handcock, P. Pattison, Recent developments inexponential random graph (p) models for social networks, Social
Networks 29 (2007) 192–215.
[141] O. Frank, D. Strauss, Markov graphs, J. Amer. Statist. Assoc. 81 (395) (1986) 832–842.
[142] D. Lusher, J. Koskinen, G. Robins, Exponential Random Graph Models for Social Networks, Cambridge University Press, Cambridge, 2013, p. 2013.
[143] J. Park, M.E.J. Newman, The statistical mechanics of networks, Phys. Rev. E 70 (2004) 066117.
[144] G. Bianconi, The entropy of randomized network ensembles, Europhys. Lett. 81 (2008) 28005.
[145] T. Squartini, D. Garlaschelli, Analytical maximum-likelihood method to detect patterns in real networks, New J. Phys. 13 (2011) 083001.
[146] K. Anand, G. Bianconi, Entropy measures for networks: toward an information theory of complex topologies, Phys. Rev. E 80 (2009) 045102.
[147] M.M.G. Bianconi, P. Pin, Assessing the relevance of node features for network structure, Proc. Natl. Acad. Sci. USA 106 (28) (2009) 11433–11438.
[148] M. Rosvall, C.T. Bergstrom, An information-theoretic framework for resolving community structure in complex networks, Proc. Natl. Acad. Sci. USA
104 (18) (2007) 7327–7331.
[149] A. Halu, S. Mukherjee, G. Bianconi, Emergence of overlap in ensembles of spatial multiplexes and statistical mechanics of spatial interacting network
ensembles, Phys. Rev. E 89 (2014) 012806.
[150] E. Lazega, P.E. Pattison, Multiplexity, generalized exchange and cooperation in organizations: a case study, Social Networks 21 (1999) 67–90.
[151] M.T. Heaney, Multiplex networks and interest group influence reputation: an exponential random graph model, Social Networks 36 (2014) 66–81.
[152] G. Bianconi, S.N. Dorogovtsev, J.F.F. Mendes, Mutually connected component of network of networks, ArXiv e-prints. arXiv:1402.0215.
[153] J. Gao, S.V. Buldyrev, S. Havlin, H.E. Stanley, Robustness of a network of networks, Phys. Rev. Lett. 107 (19) (2011) 195701.
[154] g. Bianconi, S.N. Dorogovtsev, Multiple percolation transitions in a configuration model of network of networks, ArXiv e-prints. arXiv:1402.0218.
[155] J. Gao, S.V. Buldyrev, S. Havlin, H.E. Stanley, Robustness of a network formed by n interdependent networks with a one-to-one correspondence of
dependent nodes, Phys. Rev. E 85 (6 Pt 2) (2012) 066134.
[156] J. Gao, S.V. Buldyrev, H.E. Stanley, X. Xu, S. Havlin, Percolation of a general network of networks, Phys. Rev. E 88 (2013) 062816.
[157] D.S. Callaway, M.E.J. Newman, S.H. Strogatz, D.J. Watts, Network robustness and fragility: percolation on random graphs, Phys. Rev. Lett. 85 (2000)
5468–5471.
[158] R. Cohen, K. Erez, D. ben Avraham, S. Havlin, Resilience of the internet to random breakdowns, Phys. Rev. Lett. 85 (2000) 4626–4628.
[159] R. Cohen, K. Erez, D. ben Avraham, S. Havlin, Breakdown of the internet under intentional attack, Phys. Rev. Lett. 86 (2001) 3682–3685.
[160] R. Cohen, D. ben Avraham, S. Havlin, Percolation critical exponents in scale-free networks, Phys. Rev. E 66 (2002) 036113.
[161] A. Vázquez, Y. Moreno, Resilience to damage of graphs with degree correlations, Phys. Rev. E 67 (2003) 015101.
[162] P. Crucitti, V. Latora, M. Marchiori, A. Rapisarda, Efficiency of scale-free networks: error and attack tolerance, Physica A 320 (0) (2003) 622–642.
[163] P. Crucitti, V. Latora, M. Marchiori, A. Rapisarda, Error and attack tolerance of complex networks, Physica A 340 (1–3) (2004) 388–394.
[164] A. Vespignani, Complex networks: the fragility of interdependency, Nature 464 (7291) (2010) 984–985.
[165] S.-W. Son, P. Grassberger, M. Paczuski, Percolation transitions are not always sharpened by making networks interdependent, Phys. Rev. Lett. 107
(2011) 195702.
[166] Y. Berezin, A. Bashan, S. Havlin, Comment on ‘‘percolation transitions are not always sharpened by making networks interdependent’’, Phys. Rev.
Lett. 111 (2013) 189601.
[167] S.-W. Son, P. Grassberger, M. Paczuski, Son, grassberger, and paczuski reply, Phys. Rev. Lett. 111 (2013) 189602.
[168] G.J. Baxter, S.N. Dorogovtsev, A.V. Goltsev, J.F.F. Mendes, Avalanche collapse of interdependent networks, Phys. Rev. Lett. 109 (2012) 248701.
[169] J. Shao, S.V. Buldyrev, S. Havlin, H.E. Stanley, Cascade of failures in coupled network systems with multiple support-dependence relations, Phys. Rev.
E 83 (2011) 036116.
[170] X. Huang, J. Gao, S.V. Buldyrev, S. Havlin, H.E. Stanley, Robustness of interdependent networks under targeted attack, Phys. Rev. E 83 (2011) 065101.
[171] S.-W. Son, G. Bizhani, C. Christensen, P. Grassberger, M. Paczuski, Percolation theory on interdependent networks based on epidemic spreading,
Europhys. Lett. 97 (1) (2012) 16006.
[172] M. Mezard, A. Montanari, Information, Physics and Computation, Oxford University Press, Oxford, 2009.
[173] M. Mézard, G. Parisi, The cavity method at zero temperature, J. Stat. Phys. 111 (1–2) (2003) 1–34.
[174] A.K. Hartmann, M. Weigt, Phase Transitions in Combinatorial Optimization Problems: Basics, Algorithms and Statistical Mechanics, Wiley-VCH,
Weinheim, 2005.
[175] S. Watanabe, Y. Kabashima, Cavity-based robustness analysis of interdependent networks: influences of intranetwork and internetwork degree–
degree correlations, Phys. Rev. E 89 (2014) 012808.
[176] O. Yağan, V. Gligor, Analysis of complex contagions in random multiplex networks, Phys. Rev. E 86 (2012) 036103.
[177] R. Parshani, S.V. Buldyrev, S. Havlin, Critical effect of dependency groups on the function of networks, Proc. Natl. Acad. Sci. USA 108 (3) (2011)
1007–1010.
[178] A. Bashan, R. Parshani, S. Havlin, Percolation in networks composed of connectivity and dependency links, Phys. Rev. E 83 (2011) 051127.
[179] Y. Hu, B. Ksherim, R. Cohen, S. Havlin, Percolation in interdependent and interconnected networks: abrupt change from second- to first-order
transitions, Phys. Rev. E 84 (2011) 066116.
[180] D. Zhou, J. Gao, H.E. Stanley, S. Havlin, Percolation of partially interdependent scale-free networks, Phys. Rev. E 87 (2013) 052812.
[181] Y. Hu, D. Zhou, R. Zhang, Z. Han, C. Rozenblat, S. Havlin, Percolation of interdependent networks with intersimilarity, Phys. Rev. E 88 (2013) 052805.
[182] M. Li, R.-R. Liu, C.-X. Jia, B.-H. Wang, Critical effects of overlapping of connectivity and dependence links on percolation of networks, New J. Phys. 15
(9) (2013) 093013.
[183] S.V. Buldyrev, N.W. Shere, G.A. Cwilich, Interdependent networks with identical degrees of mutually dependent nodes, Phys. Rev. E 83 (2011) 016112.
[184] R. Parshani, C. Rozenblat, D. Ietri, C. Ducruet, S. Havlin, Inter-similarity between coupled networks, Europhys. Lett. 92 (6) (2010) 68002.
[185] D. Zhou, H.E. Stanley, G. D’Agostino, A. Scala, Assortativity decreases the robustness of interdependent networks, Phys. Rev. E 86 (2012) 066103.
[186] X. Huang, S. Shao, H. Wang, S.V. Buldyrev, H.E. Stanley, S. Havlin, The robustness of interdependent clustered networks, Europhys. Lett. 101 (1) (2013)
18002.
[187] L.D. Valdez, P.A. Macri, H.E. Stanley, L.A. Braunstein, Triple point in correlated interdependent networks, Phys. Rev. E 88 (2013) 050803.
[188] W. Li, A. Bashan, S.V. Buldyrev, H.E. Stanley, S. Havlin, Cascading failures in interdependent lattice networks: the critical role of the length of
dependency links, Phys. Rev. Lett. 108 (2012) 228702. http://dx.doi.org/10.1103/PhysRevLett.108.228702.
[189] A. Bashan, Y. Berezin, S.V. Buldyrev, S. Havlin, The extreme vulnerability of interdependent spatially embedded networks, Nat. Phys. 9 (10) (2013)
667–672.
[190] M. Stippinger, J. Kertész, Enhancing resilience of interdependent networks by healing, ArXiv e-prints arXiv:1312.1993.
[191] Y. Berezin, A. Bashan, M.M. Danziger, D. Li, S. Havlin, Spatially localized attacks on interdependent networks: the existence of a finite critical attack
size, ArXiv e-prints arXiv:1310.0996.
[192] M.E.J. Newman, Mixing patterns in networks, Phys. Rev. E 67 (2003) 026126.
[193] S. Guha, D. Towsley, C. Capar, A. Swami, P. Basu, Layered percolation, ArXiv e-prints arXiv:1402.7057.
[194] K. Zhao, G. Bianconi, Percolation on interacting, antagonistic networks, J. Stat. Mech. Theory Exp. 2013 (05) (2013) P05005.
[195] N. Azimi-Tafreshi, J. Gómez-Gardeñes, S.N. Dorogovtsev, k-core percolation on multiplex networks, ArXiv e-prints arXiv:1405.1336.
[196] B. Min, K.-I. Goh, Multiple resource demands and viability in multiplex networks, Phys. Rev. E 89 (2014) 040802.
[197] G.J. Baxter, S.N. Dorogovtsev, J.F.F. Mendes, D. Cellai, Weak percolation on multiplex networks, Phys. Rev. E 89 (2014) 042801.
[198] K.-M. Lee, K.-I. Goh, I.-M. Kim, Sandpiles on multiplex networks, J. Korean Phys. Soc. 60 (4) (2012) 641–647.
[199] P. Bak, C. Tang, K. Wiesenfeld, Self-organized criticality, Phys. Rev. A 38 (1988) 364–374.
[200] D.J. Watts, A simple model of global cascades on random networks, Proc. Natl. Acad. Sci. USA 99 (9) (2002) 5766–5771.
[201] K.-M. Lee, C.D. Brummitt, K.-I. Goh, Slowed yet explosive global cascades driven by response heterogeneity in multiplex networks, ArXiv e-prints
arXiv:1403.3472.
[202] P. Blanchard, D. Volchenkov, Random Walks and Diffusions on Graphs and Databases, Springer, 2011.
[203] B. Tadic, G.J. Rodgers, Packet transport on scale-free networks, Adv. Complex Syst. (ACS) 05 (04) (2002) 445–456.
[204] J.D. Noh, H. Rieger, Random walks on complex networks, Phys. Rev. Lett. 92 (11) (2004) 118701.
[205] S.-J. Yang, Exploring complex networks by walking on them, Phys. Rev. E 71 (2005) 016107.
[206] L.F. Costa, O. Sporns, L. Antiqueira, M.G.V. Nunes, O.N. Oliveira Jr., Correlations between structure and random walk dynamics in directed complex
networks, Appl. Phys. Lett. 91 (5) (2007) 054107.
[207] S. Condamin, O. Bénichou, V. Tejedor, R. Voituriez, J. Klafter, First-passage times in complex scale-invariant media, Nature 450 (7166) (2007) 77–80.
[208] J. Gómez-Gardeñes, V. Latora, Entropy rate of diffusion processes on complex networks, Phys. Rev. E 78 (6 Pt 2) (2008) 065102.
[209] S. Valverde, R.V. Solé, Self-organized critical traffic in parallel computer networks, Physica A 312 (3) (2002) 636–648.
[210] R. Guimera, A. Díaz-Guilera, F. Vega-Redondo, A. Cabrales, A. Arenas, Optimal network topologies for local search with congestion, Phys. Rev. Lett.
89 (24) (2002) 248701.
[211] P. Echenique, J. Gómez-Gardeñes, Y. Moreno, Dynamics of jamming transitions in complex networks, Europhys. Lett. 71 (2) (2005) 325.
[212] D. De Martino, L. Dall’Asta, G. Bianconi, M. Marsili, Congestion phenomena on complex networks, Phys. Rev. E 79 (1 Pt 2) (2009) 015101.
[213] M. Fiedler, Algebraic connectivity of graphs, Czechoslovak Math. J. 23 (2) (1973) 298–305.
[214] M.E.J. Newman, Networks: An Introduction, Oxford University Press, New York, 2010.
[215] R. Cohen, S. Havlin, Complex Networks: Structure, Robustness and Function, Cambridge University Press, Cambridge, 2010.
[216] E. Estrada, N. Hatano, Communicability in complex networks, Phys. Rev. E 77 (2008) 036111.
[217] E. Estrada, N. Hatano, M. Benzi, The physics of communicability in complex networks, Phys. Rep. 514 (3) (2012) 89–119.
[218] E. Estrada, J. Gómez-Gardeñes, Communicability reveals a transition to coordinated behavior in multiplex networks, ArXiv e-prints arXiv:1312.3234.
[219] M. Marchiori, The quest for correct information on the web: Hyper search engines, Comput. Netw. ISDN Syst. 29 (8) (1997) 1225–1235.
[220] A. Arenas, A. Díaz-Guilera, R. Guimerá, Communication in networks with hierarchical branching, Phys. Rev. Lett. 86 (2001) 3196.
[221] R. Guimerá, A. Díaz-Guilera, F. Vega-Redondo, A. Cabrales, A. Arenas, Optimal network topologies for local search with congestion, Phys. Rev. Lett.
89 (2002) 248701.
[222] Z. Toroczkai, K. Bassler, Network dynamics: jamming is limited in scale-free systems, Nature 428 (2004) 716.
[223] P. Echenique, J. Gómez-Gardeñes, Y. Moreno, Improved routing strategies for internet traffic delivery, Phys. Rev. E 70 (2004) 056105.
[224] J. Zhou, G. Yan, C.-H. Lai, Efficient routing on multilayered communication networks, Europhys. Lett. 102 (2) (2013) 28002.
[225] Y. Zhuo, Y. Peng, C. Liu, Y. Liu, K. Long, Traffic dynamics on layered complex networks, Physica A 391 (2011) 2401.
[226] F. Tan, J. Wu, Y. Xia, C.K. Tse, Traffic congestion in interconnected complex networks, ArXiv e-prints arXiv:1401.0412.
[227] S.N. Dorogovtsev, A.V. Goltsev, J.F.F. Mendes, Critical phenomena in complex networks, Rev. Modern Phys. 80 (2008) 1275–1335.
[228] A.V. Goltsev, S.N. Dorogovtsev, J.G. Oliveira, J.F.F. Mendes, Localization and spreading of diseases in complex networks, Phys. Rev. Lett. 109 (2012)
128702. http://dx.doi.org/10.1103/PhysRevLett.109.128702.
[229] M. Boguñá, C. Castellano, R. Pastor-Satorras, Nature of the epidemic threshold for the susceptible-infected-susceptible dynamics in networks, Phys.
Rev. Lett. 111 (2013) 068701. http://dx.doi.org/10.1103/PhysRevLett.111.068701.
[230] R. Anderson, R. May, B. Anderson, Infectious Diseases of Humans: Dynamics and Control, Oxford University Press, Oxford, 1992.
[231] H. Hethcote, The mathematics of infectious diseases, SIAM Rev. 42 (2000) 599–653.
[232] D. Daley, J. Gani, Epidemic Modelling, Cambridge University Press, Cambridge, 1999.
[233] J. Murray, Mathematical Biology, Springer-Verlag, Berlin, 2002.
[234] J. Marro, R. Dickman, Nonequilibrium Phase Transition in Lattice Model, Cambridge University Press, Cambridge, 1999.
[235] S. Gómez, A. Arenas, J. Borge-Holthoefer, S. Meloni, Y. Moreno, Discrete-time markov chain approach to contact-based disease spreading in complex
networks, Europhys. Lett. 89 (3) (2010) 38009.
[236] F. Chung, L. Lu, V. Vu, Spectra of random graphs with given expected degrees, Proc. Natl. Acad. Sci. USA 100 (11) (2003) 6313–6318.
[237] B. Guerra, J. Gómez-Gardeñes, Annealed and mean-field formulations of disease dynamics on static and adaptive networks, Phys. Rev. E 82 (3 Pt 2)
(2010) 035101.
[238] S. Gómez, J. Gómez-Gardeñes, Y. Moreno, A. Arenas, Nonperturbative heterogeneous mean-field approach to epidemic spreading in complex
networks, Phys. Rev. E 84 (3 Pt 2) (2011) 036105.
[239] R. Pastor-Satorras, A. Vespignani, Epidemic spreading in scale-free networks, Phys. Rev. Lett. 86 (14) (2001) 3200–3203.
[240] R. Pastor-Satorras, A. Vespignani, Epidemic dynamics and endemic states in complex networks, Phys. Rev. E 63 (6 Pt 2) (2001) 066117.
[241] M. Boguñá, R. Pastor-Satorras, Epidemic spreading in correlated complex networks, Phys. Rev. E 66 (4 Pt 2) (2002) 047104.
[242] A. Saumell-Mendiola, M.A. Serrano, M. Boguna, Epidemic spreading on interconnected networks, Phys. Rev. E 86 (2 Pt 2) (2012) 026106.
[243] F. Sahneh, C. Scoglio, F. Chowdhury, Effect of coupling on the epidemic threshold in interconnected complex networks: a spectral analysis,
in: American Control Conference, ACC, 2013, pp. 2307–2312.
[244] H. Wang, Q. Li, G. D’Agostino, S. Havlin, H.E. Stanley, P. Van Mieghem, Effect of the interconnected network structure on the epidemic threshold,
Phys. Rev. E 88 (2) (2013) 022801.
[245] M. Dickison, S. Havlin, H.E. Stanley, Epidemics on interconnected networks, Phys. Rev. E 85 (6 Pt 2) (2012) 066109.
[246] R. Vida, J. Galeano, S. Cuenda, Vulnerability of overlay networks under malware spreading, ArXiv e-prints arXiv:1310.0741.
[247] B. Min, K.-I. Goh, Layer-crossing overhead and information spreading in multiplex social networks, ArXiv e-prints arXiv:1307.2967.
[248] O. Yagan, D. Qian, J. Zhang, D. Cochran, Conjoining speeds up information diffusion in overlaying social-physical networks, IEEE J. Sel. Areas Commun.
31 (6) (2013) 1038–1048.
[249] M.E. Newman, Threshold effects for two pathogens spreading on a network, Phys. Rev. Lett. 95 (10) (2005) 108701.
[250] B. Karrer, M.E. Newman, Competing epidemics on complex networks, Phys. Rev. E 84 (3 Pt 2) (2011) 036106.
[251] Y. Wang, G. Xiao, J. Liu, Dynamics of competing ideas in complex social systems, New J. Phys. 14 (1) (2012) 013015.
[252] F. Darabi Sahneh, C. Scoglio, May the best meme win!: new exploration of competitive epidemic spreading over arbitrary multi-layer networks,
ArXiv e-prints arXiv:1308.4880.
[253] X. Wei, N. Valler, B.A. Prakash, I. Neamtiu, M. Faloutsos, C. Faloutsos, Competing memes propagation on networks: a case study of composite networks,
SIGCOMM Comput. Commun. Rev. 42 (5) (2012) 5–12.
[254] X. Wei, N. Valler, B.A. Prakash, I. Neamtiu, M. Faloutsos, C. Faloutsos, Competing memes propagation on networks: a network science perspective,
IEEE J. Sel. Areas Commun. 31 (6) (2013) 1049–1060.
[255] P.V. Mieghem, J. Omic, R. Kooij, Virus spread in networks, IEEE/ACM Trans. Netw. 17 (1) (2009) 1–14.
[256] C. Granell, S. Gómez, A. Arenas, Dynamical interplay between awareness and epidemic spreading in multiplex networks, Phys. Rev. Lett. 111 (12)
(2013) 128701. http://dx.doi.org/10.1103/PhysRevLett.111.128701.
[257] C. Buono, L.G. Alvarez Zuzek, P.A. Macri, L.A. Braunstein, Epidemics in partially overlapped multiplex networks, ArXiv e-prints arXiv:1310.1939.
[258] D. Zhao, L. Li, H. Peng, Q. Luo, Y. Yang, Multiple routes transmitted epidemics on multiplex networks, Phys. Lett. A 378 (2014) 770–776. arXiv:1312.
6931.
[259] C.A. Gilligan, An epidemiological framework for disease management, Adv. Bot. Res. 38 (2002) 1–64.
[260] W.H. Organization, Global tuberculosis control: WHO report 2010, World Health Organization, 2010.
[261] R.A. Fouchier, T. Kuiken, M. Schutten, G. Van Amerongen, G.J. van Doornum, B.G. van den Hoogen, M. Peiris, W. Lim, K. Stöhr, A.D. Osterhaus, Aetiology:
Koch’s postulates fulfilled for sars virus, Nature 423 (6937) (2003) 240.
[262] M. Small, D.M. Walker, C.K. Tse, Scale-free distribution of avian influenza outbreaks, Phys. Rev. Lett. 99 (18) (2007) 188702.
[263] C. Fraser, C.A. Donnelly, S. Cauchemez, W.P. Hanage, M.D. Van Kerkhove, T.D. Hollingsworth, J. Griffin, R.F. Baggaley, H.E. Jenkins, E.J. Lyons, et al.,
Pandemic potential of a strain of influenza a (h1n1): early findings, Science 324 (5934) (2009) 1557–1561.
[264] R. Pastor-Satorras, A. Vespignani, Immunization of complex networks, Phys. Rev. E 65 (3) (2002) 036104.
[265] J. Müller, B. Schönfisch, M. Kirkilionis, Ring vaccination, J. Math. Biol. 41 (2) (2000) 143–171.
[266] R. Cohen, S. Havlin, D. Ben-Avraham, Efficient immunization strategies for computer networks and populations, Phys. Rev. Lett. 91 (24) (2003)
247901.
[267] V.M. Eguiluz, M.G. Zimmermann, Transmission of information and herd behavior: an application to financial markets, Phys. Rev. Lett. 85 (26) (2000)
5659.
[268] A. Nowak, J. Szamrej, B. Latané, From private attitude to public opinion: a dynamic theory of social impact, Psychol. Rev. 97 (3) (1990) 362.
[269] S. Funk, E. Gilad, C. Watkins, V.A. Jansen, The spread of awareness and its impact on epidemic outbreaks, Proc. Natl. Acad. Sci. USA 106 (16) (2009)
6872–6877.
[270] P.E. Fine, J.A. Clarkson, Individual versus public priorities in the determination of optimal vaccination policies, Am. J. Epidemiol. 124 (6) (1986)
1012–1020.
[271] C.T. Bauch, Imitation dynamics predict vaccinating behaviour, Proc. R. Soc. B 272 (1573) (2005) 1669–1675.
[272] C. for Disease Control, Prevention, Preventing seasonal flue., Atlanta: Centers for Disease Control and Prevention, 2008.
[273] P.-Y. Geoffard, T. Philipson, Disease eradication: private versus public vaccination, Am. Econ. Rev. 87 (1) (1997) 222–230.
[274] Y. Ibuka, M. Li, J. Vietri, G.B. Chapman, A.P. Galvani, Free-riding behavior in vaccination decisions: an experimental study, PLoS ONE 9 (1) (2014)
e87164.
[275] P.J. Francis, Dynamic epidemiology and the market for vaccinations, J. Public Econ. 63 (3) (1997) 383–406.
[276] C.T. Bauch, A.P. Galvani, D.J. Earn, Group interest versus self-interest in smallpox vaccination policy, Proc. Natl. Acad. Sci. USA 100 (18) (2003)
10564–10567.
[277] C.T. Bauch, D.J. Earn, Vaccination and the theory of games, Proc. Natl. Acad. Sci. USA 101 (36) (2004) 13391–13394.
[278] H. Zhang, J. Zhang, C. Zhou, M. Small, B. Wang, Hub nodes inhibit the outbreak of epidemic under voluntary vaccination, New J. Phys. 12 (2) (2010)
023015.
[279] A. Cardillo, C. Reyes-Suárez, F. Naranjo, J. Gómez-Gardeñes, Evolutionary vaccination dilemma in complex networks, Phys. Rev. E 88 (3) (2013)
032803.
[280] F. Fu, D.I. Rosenbloom, L. Wang, M.A. Nowak, Imitation dynamics of vaccination behaviour on social networks, Proc. R. Soc. B 278 (1702) (2011)
42–49.
[281] X.-T. Liu, Z.-X. Wu, L. Zhang, Impact of committed individuals on vaccination behavior, Phys. Rev. E 86 (5) (2012) 051132.
[282] L. Mao, Y. Yang, Coupling infectious diseases, human preventive behavior, and networks—a conceptual framework for epidemic modeling, Soc. Sci.
Med. 74 (2) (2012) 167–175.
[283] K. Stöhr, M. Esveld, Will vaccines be available for the next influenza pandemic? Science 306 (5705) (2004) 2195–2196.
[284] L. Mao, Modeling triple-diffusions of infectious diseases, information, and preventive behaviors through a metropolitan social network—an agent-
based simulation, Appl. Geogr. 50 (2014) 31–39.
[285] C.T. Bauch, A.P. Galvani, Social factors in epidemiology, Science 342 (6154) (2013) 47–49.
[286] S. Redner, A Guide to First-Passage Processes, Cambridge University Press, 2001.
[287] C. Castellano, S. Fortunato, V. Loreto, Statistical physics of social dynamics, Rev. Modern Phys. 81 (2) (2009) 591.
[288] R. Hegselmann, U. Krause, Opinion dynamics and bounded confidence models, analysis, and simulation, J. Artif. Soc. Soc. Simul. 5 (2002) 3.
[289] J.H. Fowler, C.T. Dawes, Two genes predict voter turnout, J. Polit. 70 (03) (2008) 579–594.
[290] K. Sznajd-Weron, J. Sznajd, Opinion evolution in closed community, Internat. J. Modern Phys. C 11 (06) (2000) 1157–1165.
[291] P. Krapivsky, S. Redner, Dynamics of majority rule in two-state interacting spin systems, Phys. Rev. Lett. 90 (23) (2003) 238701.
[292] Z.-X. Wu, P. Holme, Majority-vote model on hyperbolic lattices, Phys. Rev. E 81 (1) (2010) 011133.
[293] B. Kozma, A. Barrat, Consensus formation on adaptive networks, Phys. Rev. E 77 (1) (2008) 016102.
[294] R.A. Holley, T.M. Liggett, Ergodic theorems for weakly interacting infinite systems and the voter model, Ann. Probab. (1975) 643–663.
[295] K. Suchecki, V.M. Eguíluz, M. San Miguel, Voter model dynamics in complex networks: role of dimensionality, disorder, and degree distribution,
Phys. Rev. E 72 (3) (2005) 036132.
[296] V. Sood, S. Redner, Voter model on heterogeneous graphs, Phys. Rev. Lett. 94 (17) (2005) 178701.
[297] J. Shao, S. Havlin, H.E. Stanley, Dynamic opinion model and invasion percolation, Phys. Rev. Lett. 103 (1) (2009) 018701.
[298] R.J. Glauber, Time-dependent statistics of the ising model, J. Math. Phys. 4 (2) (2004) 294–307.
[299] I. Dornic, H. Chaté, J. Chave, H. Hinrichseh, Critical coarsening without surface tension: the universality class of the voter model, Phys. Rev. Lett. 87
(4) (2001) 045701.
[300] R. Lambiotte, S. Redner, Dynamics of vacillating voters, J. Stat. Mech. Theory Exp. 2007 (10) (2007) L10001.
[301] J. Molofsky, R. Durrett, J. Dushoff, D. Griffeath, S. Levin, Local frequency dependence and global coexistence, Theor. Popul. Biol. 55 (3) (1999) 270–282.
[302] N. Masuda, N. Gibert, S. Redner, Heterogeneous voter models, Phys. Rev. E 82 (1) (2010) 010103.
[303] F. Vazquez, P.L. Krapivsky, S. Redner, Constrained opinion dynamics: freezing and slow evolution, J. Phys. A: Math. Gen. 36 (3) (2003) L61.
[304] F. Slanina, H. Lavicka, Analytical results for the sznajd model of opinion formation, Eur. Phys. J. B 35 (2) (2003) 279–288.
[305] C. Castellano, V. Loreto, A. Barrat, F. Cecconi, D. Parisi, Comparison of voter and glauber ordering dynamics on networks, Phys. Rev. E 71 (6) (2005)
066107.
[306] K. Suchecki, V.M. Eguiluz, M. San Miguel, Conservation laws for the voter model in complex networks, Europhys. Lett. 69 (2) (2005) 228.
[307] V. Sood, T. Antal, S. Redner, Voter models on heterogeneous networks, Phys. Rev. E 77 (4) (2008) 041121.
[308] C. Castellano, D. Vilone, A. Vespignani, Incomplete ordering of the voter model on small-world networks, Europhys. Lett. 63 (1) (2003) 153.
[309] D. Vilone, C. Castellano, Solution of voter model dynamics on annealed small-world networks, Phys. Rev. E 69 (1) (2004) 016109.
[310] G. Zschaler, G.A. Böhme, M. Seißinger, C. Huepe, T. Gross, Early fragmentation in the adaptive voter model on directed networks, Phys. Rev. E 85 (4)
(2012) 046107.
[311] B. Latane, The psychology of social impact, Am. Psychol. 36 (4) (1981) 343.
[312] S. Galam, Heterogeneous beliefs, segregation, and extremism in the making of public opinions, Phys. Rev. E 71 (4) (2005) 046123.
[313] Z. Wang, Y. Liu, L. Wang, Y. Zhang, Freezing period strongly impacts the emergence of a global consensus in the voter model, Sci. Rep. 4 (2014) 3597.
[314] H.-U. Stark, C.J. Tessone, F. Schweitzer, Decelerating microdynamics can accelerate macrodynamics in the voter model, Phys. Rev. Lett. 101 (1) (2008)
018701.
[315] H.-U. Stark, C.J. Tessone, F. Schweitzer, Slower is faster: fostering consensus formation by heterogeneous inertia, Adv. Complex Syst. 11 (04) (2008)
551–563.
[316] F. Vazquez, V.M. Eguíluz, M. San Miguel, Generic absorbing transition in coevolution dynamics, Phys. Rev. Lett. 100 (10) (2008) 108702.
[317] M. Mobilia, Does a single zealot affect an infinite group of voters? Phys. Rev. Lett. 91 (2) (2003) 028701.
[318] M. Mobilia, A. Petersen, S. Redner, On the role of zealotry in the voter model, J. Stat. Mech. Theory Exp. 2007 (08) (2007) P08029.
[319] M. Mobilia, I.T. Georgiev, Voting and catalytic processes with inhomogeneities, Phys. Rev. E 71 (4) (2005) 046102.
[320] T. Antal, S. Redner, V. Sood, Evolutionary dynamics on degree-heterogeneous graphs, Phys. Rev. Lett. 96 (18) (2006) 188104.
[321] T. Zillio, I. Volkov, J.R. Banavar, S.P. Hubbell, A. Maritan, Spatial scaling in model plant communities, Phys. Rev. Lett. 95 (2005) 098101.
[322] N. Masuda, Voter model on the two-clique graph, arXiv preprint arXiv:1403.4763.
[323] M. Diakonova, M. San Miguel, V.M. Eguiluz, Absorbing and shattered fragmentation transitions in multilayer coevolution, ArXiv e-prints
arXiv:1403.4534.
[324] A. Halu, K. Zhao, A. Baronchelli, G. Bianconi, Connect and win: the role of social networks in political elections, Europhys. Lett. 102 (1) (2013) 16002.
[325] J. Xie, S. Sreenivasan, G. Korniss, W. Zhang, C. Lim, B.K. Szymanski, Social consensus through the influence of committed minorities, Phys. Rev. E 84
(1) (2011) 011130.
[326] G. Szabó, G. Fáth, Evolutionary games on graphs, Phys. Rep. 446 (4–6) (2007) 97–216.
[327] C.P. Roca, J.A. Cuesta, A. Sánchez, Evolutionary game theory: temporal and spatial effects beyond replicator dynamics, Phys. Life Rev. 6 (4) (2009)
208–249.
[328] M. Perc, J. Gómez-Gardeñes, A. Szolnoki, L.M. Floría, Y. Moreno, Evolutionary dynamics of group interactions on structured populations: a review,
J. R. Soc. Interface 6 (2013) 20120997.
[329] M. Nowak, Evolutionary Dynamics: Exploring the Equations of Life, Belknap Press of Hardvard University Press, Cambridge, Massachusets, 2006.
[330] M. Nowak, Five rules for the evolution of cooperation, Science 8 (5805) (2006) 1560–1563.
[331] J. Maynard-Smith, E. Szathmary, The Major Transitions in Evolution, Freeman, Oxford, 1995.
[332] P. Kappeler, C. van Schaik (Eds.), Cooperation in Primates and Humans: Mechanisms and Evolution, Springer-Verlag, Berlin-Heidelberg, 2006.
[333] E. Wilson, Sociobiology, Belknap Press of Harvard University Press, Cambridge, Massachusets, 2000.
[334] J. Henrich, R. Boyd, S. Bowles, C. Camere, E. Fehr, H.G. (Eds.), Foundations of Human Sociality: Economic Experiments and Ethnographic Evidence
from Fifteen Small-Scale Societies, Oxford University Press, Oxford, 2004.
[335] A. Rapoport, M. Guyer, A taxonomy of 2 × 2 games, Gen. Syst.: Yearb. Soc. Gen. Syst. Res. 11 (1996) 203–214.
[336] R. Axelrod, The Evolution of Cooperation, Basic Books, New York, 1984.
[337] M. Macy, A. Flache, Learning dynamics in social dilemmas, Proc. Natl. Acad. Sci. USA 99 (2002) 7229–7236.
[338] J. Hofbauer, K. Sigmund, Evolutionary Games and Population Dynamics, Cambridge University Press, Cambridge, 1998.
[339] K. Sigmund, The Calculus of Selfishness, Princeton University Press, Princeton, New Yersey, 2010.
[340] M.A. Nowak, R.M. May, Evolutionary games and spatial chaos, Nature 359 (6398) (1992) 826–829.
[341] F.C. Santos, J.M. Pacheco, Scale-free networks provide a unifying framework for the emergence of cooperation, Phys. Rev. Lett. 95 (2005) 098104.
[342] J. Gómez-Gardeñes, M. Campillo, L.M. Floria, Y. Moreno, Dynamical organization of cooperation in complex topologies, Phys. Rev. Lett. 98 (10) (2007)
108103.
[343] J. Kagel, A. Roth, The Handbook of Experimental Economics, Princeton University Press, Cambridge, Massachusets, 1995.
[344] H. Brandt, C. Hauert, K. Sigmund, Punishment and reputation in spatial public goods games, Proc. R. Soc. B 270 (1519) (2003) 1099–1104.
[345] C. Hauert, G. Szabó, Prisoner’s dilemma and public goods games in different geometries: compulsory versus voluntary interactions, Complexity 8 (4)
(2003) 31–38.
[346] F.C. Santos, M.D. Santos, J.M. Pacheco, Social diversity promotes the emergence of cooperation in public goods games, Nature 454 (7201) (2008)
213–216.
[347] J. Gómez-Gardeñes, M. Romance, R. Criado, D. Vilone, A. Sánchez, Evolutionary games defined at the network mesoscale: the public goods game,
Chaos 21 (1) (2011) 016113.
[348] J. Gómez-Gardeñes, D. Vilone, A. Sánchez, Disentangling social and group heterogeneities: public goods games on complex networks, Europhys. Lett.
95 (6) (2011) 68003.
[349] J. Gómez-Gardeñes, C. Gracia-Lazaro, L.M. Floria, Y. Moreno, Evolutionary dynamics on interdependent populations, Phys. Rev. E 86 (5 Pt 2) (2012)
056113.
[350] L.-L. Jiang, M. Perc, Spreading of cooperative behaviour across interdependent groups, Sci. Rep. 3 (2013) 2483.
[351] M. Santos, S. Dorogovtsev, J. Mendes, Biased imitation in coupled evolutionary games in interdependent networks, Sci. Rep. 4 (2014) 4436.
[352] J. Gómez-Gardeñes, I. Reinares, A. Arenas, M. Floria, Evolution of cooperation in multiplex networks, Sci. Rep. 2 (2012) 620.
[353] Z. Wang, L. Wang, M. Perc, Degree mixing in multilayer networks impedes the evolution of cooperation, Phys. Rev. E 89 (2014) 052813.
[354] Q. Jin, C.-y. Xia, L. Wang, Z. Wang, Spontaneous symmetry breaking in interdependent networked game, Sci. Rep. 4 (2014) 4095.
[355] Z. Wang, A. Szolnoki, M. Perc, Evolution of public cooperation on interdependent networks: the impact of biased utility functions, Europhys. Lett. 97
(4) (2012) 48001.
[356] Z. Wang, A. Szolnoki, M. Perc, Interdependent network reciprocity in evolutionary games, Sci. Rep. 3 (2013) 1183.
[357] B. Wang, X. Chen, L. Wang, Probabilistic interconnection between interdependent networks promotes cooperation in the public goods game, J. Stat.
Mech. Theory Exp. 2012 (11) (2012) P11017.
[358] A. Szolnoki, M. Perc, Information sharing promotes prosocial behaviour, New J. Phys. 15 (5) (2013) 053010.
[359] Z. Wang, A. Szolnoki, M. Perc, Optimal interdependence between networks for the evolution of cooperation, Sci. Rep. 3 (2013) 2470.
[360] M. Perc, Z. Wang, Heterogeneous aspirations promote cooperation in the prisoner’s dilemma game, PLoS ONE 5 (12) (2010) e15117.
[361] M. Perc, A. Szolnoki, Coevolutionary games—a mini review, BioSystems 99 (2) (2010) 109–125.
[362] Z. Wang, A. Szolnoki, M. Perc, Rewarding evolutionary fitness with links between populations promotes cooperation, J. Theor. Biol. 349 (2014) 50–56.
[363] G. Szabó, C. Tőke, Evolutionary prisoner’s dilemma game on a square lattice, Phys. Rev. E 58 (1998) 69–73. http://dx.doi.org/10.1103/PhysRevE.58.69.
[364] Z. Wang, A. Szolnoki, M. Perc, Self-organization towards optimally interdependent networks by means of coevolution, New J. Phys. 16 (3) (2014)
033041.
[365] F. Varela, J.-P. Lachaux, E. Rodriguez, J. Martinerie, The brainweb: phase synchronization and large-scale integration, Nat. Rev. Neurosci. 2 (4) (2001)
229–239.
[366] E.L. Berlow, Strong effects of weak interactions in ecological communities, Nature 398 (6725) (1999) 330–334.
[367] S. Boccaletti, J. Kurths, G. Osipov, D. Valladares, C. Zhou, The synchronization of chaotic systems, Phys. Rep. 366 (1–2) (2002) 1–101.
[368] L.F. Lago-Fernández, R. Huerta, F. Corbacho, J.A. Sigüenza, Fast response and temporal coherent oscillations in small-world networks, Phys. Rev. Lett.
84 (2000) 2758–2761.
[369] T. Nishikawa, A.E. Motter, Y.-C. Lai, F.C. Hoppensteadt, Heterogeneity in oscillator networks: are smaller worlds easier to synchronize? Phys. Rev.
Lett. 91 (2003) 014101.
[370] L.M. Pecora, T.L. Carroll, Master stability functions for synchronized coupled systems, Phys. Rev. Lett. 80 (1998) 2109–2112.
[371] M. Barahona, L.M. Pecora, Synchronization in small-world systems, Phys. Rev. Lett. 89 (2002) 054101.
[372] A.E. Motter, C. Zhou, J. Kurths, Enhancing complex-network synchronization, Europhys. Lett. 69 (3) (2005) 334.
[373] D.-U. Hwang, M. Chavez, A. Amann, S. Boccaletti, Synchronization in complex networks with age ordering, Phys. Rev. Lett. 94 (2005) 138701.
[374] M. Chavez, D.-U. Hwang, A. Amann, H.G.E. Hentschel, S. Boccaletti, Synchronization is enhanced in weighted complex networks, Phys. Rev. Lett. 94
(2005) 218701.
[375] F. Li, X. Lu, Complete synchronization of temporal boolean networks, Neural Netw. 44 (0) (2013) 72–77.
[376] J. Yao, H.O. Wang, Z.-H. Guan, W. Xu, Passive stability and synchronization of complex spatio-temporal switching networks with time delays,
Automatica 45 (7) (2009) 1721–1728.
[377] N. Masuda, K. Klemm, V.M. Eguíluz, Temporal networks: slowing down diffusion by long lasting interactions, Phys. Rev. Lett. 111 (2013) 188701.
[378] I.V. Belykh, V.N. Belykh, M. Hasler, Blinking model and synchronization in small-world networks with a time-varying coupling, Physica D 195 (1–2)
(2004) 188–206.
[379] D. Stilwell, E. Bollt, D. Roberson, Sufficient conditions for fast switching synchronization in time-varying network topologies, SIAM J. Appl. Dyn. Syst.
5 (1) (2006) 140–156.
[380] J.-D.J. Han, N. Bertin, T. Hao, D.S. Goldberg, G.F. Berriz, D. Zhang, Lan V. Dupuy, A.J.M. Walhout, M.E. Cusick, F.P. Roth, M. Vidal, Evidence for dynamically
organized modularity in the yeast protein–protein interaction network, Nature 430 (6995) (2004) 88–93.
[381] P.W. Anderson, K. Arrow, D. Pines, The Economy as an Evolving Complex System, Addison-Wesley Co, Redwood City, CA, 1988.
[382] S. Geršgorin, Über die abgrenzung der eigenwerte einer matrix, Bull. Acad. Sci. l’URSS. Classe Sci. Math. Et Na 6 (1931) 749–754.
[383] S. Boccaletti, D.-U. Hwang, M. Chavez, A. Amann, J. Kurths, L.M. Pecora, Synchronization in dynamical networks: evolution along commutative graphs,
Phys. Rev. E 74 (2006) 016102.
[384] J.D. Noh, H. Rieger, Stability of shortest paths in complex networks with random edge weights, Phys. Rev. E 66 (2002) 066127.
[385] A.L. Barabasi, R. Albert, Emergence of scaling in random networks, Science 286 (5439) (1999) 509–512.
[386] R.E. Amritkar, C.-K. Hu, Synchronized state of coupled dynamics on time-varying networks, Chaos 16 (1) (2006) 015117.
[387] D. Irving, F. Sorrentino, Synchronization of dynamical hypernetworks: dimensionality reduction through simultaneous block-diagonalization of
matrices, Phys. Rev. E 86 (2012) 056102.
[388] F. Sorrentino, Synchronization of hypernetworks of coupled dynamical systems, New J. Phys. 14 (3) (2012) 033035.
[389] J. Aguirre, R. Sevilla-Escoboza, R. Gutiérrez, D. Papo, M. Buldú, J., Synchronization of interconnected networks: the role of connector nodes, Phys. Rev.
Lett. 112 (2014) 248701. http://dx.doi.org/10.1103/PhysRevLett.112.248701.
[390] J. Martin-Hernandez, H. Wang, P. Van Mieghem, G. D’Agostino, On synchronization of interdependent networks, ArXiv e-prints arXiv:1304.4731.
[391] S.-J. Wang, X.-J. Xu, Z.-X. Wu, Y.-H. Wang, Effects of degree distribution in mutual synchronization of neural networks, Phys. Rev. E 74 (2006) 041915.
[392] L. Huang, K. Park, Y.-C. Lai, L. Yang, K. Yang, Abnormal synchronization in complex clustered networks, Phys. Rev. Lett. 97 (2006) 164101.
[393] A.A. Rad, I. Sendiña Nadal, D. Papo, M. Zanin, J.M. Buldú, F. del Pozo, S. Boccaletti, Topological measure locating the effective crossover between
segregation and integration in a modular network, Phys. Rev. Lett. 108 (2012) 228701.
[394] E. Barreto, B. Hunt, E. Ott, P. So, Synchronization in networks of networks: the onset of coherent collective behavior in systems of interacting
populations of heterogeneous oscillators, Phys. Rev. E 77 (2008) 036107.
[395] V.H.P. Louzada, N.A.M. Araújo, J.S. Andrade Jr., H.J. Herrmann, Breathing synchronization in interconnected networks, Sci. Rep. 3.
[396] A. Bogojeska, S. Filiposka, I. Mishkovski, L. Kocarev, On opinion formation and synchronization in multiplex networks, in: Telecommunications Forum,
TELFOR, 2013 21st, 2013, pp. 172–175.
[397] C. Li, W. Sun, J. Kurths, Synchronization between two coupled complex networks, Phys. Rev. E 76 (2007) 046204.
[398] Z. Wang, C. Cai, J. Wang, H. Zhang, Adaptive fault estimation of coupling connections for synchronization of complex interconnected networks,
in: C. Guo, Z.-G. Hou, Z. Zeng (Eds.), Advances in Neural Networks-ISNN 2013, in: Lecture Notes in Computer Science, vol. 7952, Springer, 2013,
pp. 455–462.
[399] Y. Shang, M. Chen, J. Kurths, Generalized synchronization of complex networks, Phys. Rev. E 80 (2009) 027201.
[400] X. Wu, W.X. Zheng, J. Zhou, Generalized outer synchronization between complex dynamical networks, Chaos 19 (1) (2009).
[401] Y. Wu, C. Li, Y. Wu, J. Kurths, Generalized synchronization between two different complex networks, Commun. Nonlinear Sci. Numer. Simul. 17 (1)
(2012) 349–355.
[402] Q. Xu, S. Zhuang, D. Hu, Y. Zeng, J. Xiao, Generalized mutual synchronization between two controlled interdependent networks, Abstr. Appl. Anal.
(2014).
[403] X. Mao, Stability switches, bifurcation, and multi-stability of coupled networks with time delays, Appl. Math. Comput. 218 (11) (2012) 6263–6274.
[404] Q.-C. Pham, J.-J. Slotine, Stable concurrent synchronization in dynamic system networks, Neural networks 20 (1) (2007) 62–77.
[405] F. Sorrentino, E. Ott, Network synchronization of groups, Phys. Rev. E 76 (5) (2007) 056114.
[406] W. Sun, R. Wang, W. Wang, J. Cao, Analyzing inner and outer synchronization between two coupled discrete-time networks with time delays, Cogn.
Neurodynamics 4 (3) (2010) 225–231.
[407] Y.-Q. Gu, C. Shao, X.-C. Fu, Complete synchronization and stability of star-shaped complex networks, Chaos Solitons Fractals 28 (2) (2006) 480–488.
[408] M. Liu, Y. Shao, X. Fu, Complete synchronization on multi-layer center dynamical networks, Chaos Solitons Fractals 41 (5) (2009) 2584–2591.
[409] Y.-L. Xi, Z.-Y. Wu, X.-C. Fu, Dynamical synchronization and stability of complex networks with multi-layer centers, Chaos Solitons Fractals 40 (2)
(2009) 635–641.
[410] A.D. Reyes, Synchrony-dependent propagation of firing rate in iteratively constructed networks in vitro, Nat. Neurosci. 6 (6) (2003) 593–599.
[411] T. Nowotny, R. Huerta, Explaining synchrony in feed-forward networks: are mcculloch-pitts neurons good enough? Biol. Cybernet. 89 (4) (2003)
237–241.
[412] M. Yi, L. Yang, Propagation of firing rate by synchronization and coherence of firing pattern in a feed-forward multilayer neural network, Phys. Rev.
E 81 (2010) 061924.
[413] E. Cohen, M. Rosenbluh, I. Kanter, Phase transition in crowd synchrony of delay-coupled multilayer laser networks, Opt. Express 20 (18) (2012)
19683–19689.
[414] R. Gutiérrez, I. Sendiña Nadal, M. Zanin, D. Papo, S. Boccaletti, Targeting the dynamics of complex networks, Sci. Rep. 2 (2012) 396.
[415] L.M. Pecora, T.L. Carroll, Synchronization in chaotic systems, Phys. Rev. Lett. 64 (1990) 821–824.
[416] L. Kocarev, U. Parlitz, Generalized synchronization, predictability, and equivalence of unidirectionally coupled dynamical systems, Phys. Rev. Lett.
76 (1996) 1816–1819.
[417] P. Erdős, A. Rènyi, On random graphs, I, Publicationes Mathematicae (Debrecen) 6 (1959) 290–297.
[418] L.d.F. Costa, O.N. Oliveira Jr, G. Travieso, F.A. Rodrigues, P.R. Villas Boas, L. Antiqueira, M.P. Viana, L.E. Correa Rocha, Analyzing and modeling real-world
phenomena with complex networks: a survey of applications, Adv. Phys. 60 (3) (2011) 329–412.
[419] M. Szell, S. Thurner, Measuring social dynamics in a massive multiplayer online game, Social Networks 32 (2010) 313–319.
[420] M. Szell, S. Thurner, How women organize social networks different from men, Sci. Rep. 3.
[421] P. Klimek, S. Thurner, Triadic closure dynamics drives scaling laws in social multiplex networks, New J. Phys. 15 (6) (2013) 063008.
[422] B. Corominas-Murtra, B. Fuchs, S. Thurner, Detection of the elite structure in a virtual multiplex social system by means of a generalized $K$-core,
ArXiv e-prints arXiv:1309.6740.
[423] E.-A. Horvát, K.A. Zweig, One-mode projection of multiplex bipartite graphs, in: ASONAM, IEEE Computer Society, 2012, pp. 599–606.
[424] E.-A. Horvát, K. Zweig, A fixed degree sequence model for the one-mode projection of multiplex bipartite graphs, Soc. Netw. Anal. Min. 3 (4) (2013)
1209–1224.
[425] K. Musiał, P. Kazienko, T. Kajdanowicz, Multirelational social networks in multimedia sharing systems, in: Knowledge Processing and Reasoning for
Information Society, Academic Publishing House EXIT, Warsaw, 2008, pp. 275–2982.
[426] K. Lewis, J. Kaufman, M. Gonzalez, A. Wimmer, N. Christakis, Tastes, ties, and time: a new social network dataset using facebook.com, Social Networks
30 (4) (2008) 330–342.
[427] A. Hashmi, F. Zaidi, A. Sallaberry, T. Mehmood, Are all social networks structurally similar? in: Advances in Social Networks Analysis and Mining,
ASONAM, 2012 IEEE/ACM International Conference on, IEEE, 2012, pp. 310–314.
[428] K. Lewis, M. Gonzalez, J. Kaufman, Social selection and peer influence in an online social network, Proc. Natl. Acad. Sci. USA.
[429] L. Tang, X. Wang, H. Liu, Community detection via heterogeneous interaction analysis, Data Min. Knowl. Discov. 25 (1) (2012) 1–33.
[430] A. Ansari, O. Koenigsberg, F. Stahl, Modeling multiple relationships in social networks, J. Marketing Res. 48 (4) (2011) 713–728.
[431] M. Magnani, L. Rossi, Multi-Stratum Networks: toward a unified model of on-line identities, ArXiv e-prints arXiv:1211.0169.
[432] C. Zhong, M. Salehi, S. Shah, M. Cobzarenco, N. Sastry, M. Cha, Social bootstrapping: how pinterest and Last.fm social communities benefit by
borrowing links from Facebook, ArXiv e-prints arXiv:1402.6500.
[433] G.A. Barnett, H.W. Park, K. Jiang, C. Tang, I.F. Aguillo, A multi-level network analysis of web-citations among the world’s universities, Scientometrics
99 (1) (2014) 5–26.
[434] Z. Wu, W. Yin, J. Cao, G. Xu, A. Cuzzocrea, Community detection in multi-relational social networks, in: X. Lin, Y. Manolopoulos, D. Srivastava,
G. Huang (Eds.), Web Information Systems Engineering WISE 2013, in: Lecture Notes in Computer Science, vol. 8181, Springer Berlin Heidelberg,
2013, pp. 43–56. http://dx.doi.org/10.1007/978-3-642-41154-0_4.
[435] M.K.-P. Ng, X. Li, Y. Ye, Multirank: co-ranking for objects and relations in multi-relational data, in: Proc. of the 17th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, KDD ’11, ACM, New York, NY, USA, 2011, pp. 1217–1225.
[436] X. Li, M.K. Ng, Y. Ye, Har: Hub, authority and relevance scores in multi-relational data for query search., in: SDM, SIAM/Omnipress, 2012, pp. 141–152.
[437] P. Bródka, P. Stawiak, P. Kazienko, Shortest path discovery in the multi-layered social network., in: ASONAM, IEEE Computer Society, 2011,
pp. 497–501.
[438] B. Boden, S. Günnemann, H. Hoffmann, T. Seidl, Rmics: a robust approach for mining coherent subgraphs in edge-labeled multi-layer graphs,
in: A. Szalay, T. Budavari, M. Balazinska, A. Meliou, A. Sacan (Eds.), SSDBM, ACM, 2013, p. 23.
[439] J. Sun, D. Tao, C. Faloutsos, Beyond streams and graphs: dynamic tensor analysis, in: Proc. of the 12th ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, KDD ’06, ACM, New York, NY, USA, 2006, pp. 374–383.
[440] A. Walker, The fall and rise of the scottish community, Soc. Sci. Directory 2 (3).
[441] M.T. Heaney, Multiplex networks and interest group influence reputation: an exponential random graph model, Social Networks 36 (2014) 66–81.
[442] D.C. Duling, Paul’s aegean network: the strength of strong ties, Biblical Theology Bull.: J. Bible Culture 43 (3) (2013) 135–154.
[443] M. Zignani, C. Quadri, S. Gaitto, G.P. Rossi, Exploiting all phone media? a multidimensional network analysis of phone users’ sociality, arXiv preprint
arXiv:1401.3126.
[444] A. Sen, A. Mazumder, J. Banerjee, A. Das, R. Compton, Identification of K most vulnerable nodes in multi-layered network using a new model of
interdependency, arXiv preprint arXiv:1401.1783.
[445] J.-F. Castet, J.H. Saleh, Interdependent multi-layer networks: modeling and survivability analysis with applications to space-based networks, PLoS
ONE 8 (4) (2013) e60402.
[446] P. Kaluza, A. Kölzsch, M.T. Gastner, B. Blasius, The complex network of global cargo ship movements, J. R. Soc. Interface 7 (48) (2010) 1093–1103.
[447] M. Hauge, L. Landmark, P. Lubkowski, M. Amanowicz, K. Maslanka, Selected issues of qos provision in heterogenous military networks, Int. J. Electron.
Telecommun. 60 (1) (2014) 7–13.
[448] M.C. Mahutga, The persistence of structural inequality? a network analysis of international trade, 1965–2000, Soc. Forces 84 (4) (2006) 1863–1889.
[449] C. Ducruet, Network diversity and maritime flows, J. Transport Geography 30 (2013) 77–88.
[450] L. Bargigli, G. Di Iasio, L. Infante, F. Lillo, F. Pierobon, The multiplex structure of interbank networks, arXiv preprint arXiv:1311.4798.
[451] T.L. Amburgey, A. Al-Laham, D. Tzabbar, B. Aharonson, The structural evolution of multiplex organizational networks: research and commerce in
biotechnology, Adv. Strategic Manag. 25 (2008) 171–209.
[452] S. Lee, P. Monge, The coevolution of multiplex communication networks in organizational communities, J. Commun. 61 (4) (2011) 758–779.
[453] F. Bonacina, M. D’Errico, E. Moretto, S. Stefani, A. Torriero, A multiple network approach to corporate governance, arXiv preprint arXiv:1401.4387.
[454] D. Vandamme, W. Fitzmaurice, B. Kholodenko, W. Kolch, Systems medicine: helping us understand the complexity of disease, QJM 106 (10) (2013)
891–895.
[455] L. Hood, Systems biology and p4 medicine: past, present, and future, RMMJ 4 (2) (2013) e0012.
[456] T. Michoel, B. Nachtergaele, Alignment and integration of complex networks by hypergraph-based spectral clustering, Phys. Rev. E 86 (2012) 056111.
[457] V. Nicosia, V. Latora, Measuring and modelling correlations in multiplex networks, arXiv preprint arXiv:1403.1546.
[458] W. Li, C.-C. Liu, T.Z. H., Li, M.S. Waterman, X.J. Zhou, Integrative analysis of many weighted co-expression networks using tensor computation, PLoS
Comput. Biol. 7 (6).
[459] W. Li, C. Dai, C.-C. Liu, X.J. Zhou, Algorithm to identify frequent coupled modules from two-layered network series: application to study transcription
and splicing coupling, J. Comput. Biol. 19 (6) (2012) 710–730.
[460] A. Feng, Z. Gong, Q. Wang, G. Feng, Three-dimensional air–sea interactions investigated with bilayer networks, Theor. Appl. Climatol. 109 (3–4)
(2012) 635–643.
[461] A. Mouhri, N. Flipo, F. Rejiba, C. de Fouquet, L. Bodet, G. Tallec, V. Durand, A. Jost, P. Ansart, P. Goblet, Designing a multiscale experimental sampling
system for quantifying stream-aquifer water exchanges in a multi-layer aquifer system, Geophys. Res. Abstracts, 15.
[462] O. Varol, F. Menczer, Connecting dream networks across cultures, in: Proc. of the Companion Publication of the 23rd International Conference on
World Wide Web Companion, WWW Companion ’14, 2014, pp. 1267–1272.
[463] E. Goffman, Frame Analysis: An Essay on the Organization of Experience, Harvard University Press, 1974.
[464] S.F. Nadel, M. Fortes, The Theory of Social Structure, Cohen & West, 1957.
[465] H.C. White, S.A. Boorman, R.L. Breiger, Stability of shortest paths in complex networks with random edge weights, Am. J. Sociol. 81 (1976) 730–780.
[466] B.C. Griffith, V.L. Maier, A.J. Miller, Describing communication networks through the use of matrix-based measures, unpublished paper, Graduate
School of Library Science, Drexel University.
[467] R.L. Breiger, P.E. Pattison, Cumulated social roles: the duality of persons and their algebras, Social Networks 8 (1986) 215–256.
[468] J.F. Padgett, C.K. Ansell, Robust action and the rise of the medici, Am. J. Sociol. 98 (1993) 1259–1319.
[469] K. Knipscheer, P. Dykstra, J. de Jong Gierveld, T. van Tilburg, Living Arrangements and Social Networks as Interlocking Mediating Structures, VU
University Press, Amsterdam, 1995, pp. 1–14. (Chapter 1).
[470] E. Castronova, Synthetic Worlds: The Business and Culture of Online Games, University of Chicago Press, 2005.
[471] R.I. Dunbar, Coevolution of neocortical size, group size and language in humans, Behavioral Brain Sci. 16 (04) (1993) 681–694.
[472] L. Lü, M. Medo, C.H. Yeung, Y.-C. Zhang, Z.-K. Zhang, T. Zhou, Recommender systems, Phys. Rep. 519 (1) (2012) 1–49.
[473] J. Bennett, S. Lanning, The netflix prize, in: Proc. of the KDD Cup Workshop 2007, ACM, New York, 2007, pp. 3–6.
[474] K. Shen, L. Wu, Folksonomy as a Complex Network, eprint arXiv:cs/0509072 arXiv:cs/0509072.
[475] A. Mislove, H.S. Koppula, K.P. Gummadi, P. Druschel, B. Bhattacharjee, Growth of the flickr social network, in: Proc. of the First Workshop on Online
Social Networks, WOSN ’08, ACM, New York, NY, USA, 2008, pp. 25–30.
[476] A. Nazir, S. Raza, C.-N. Chuah, Unveiling facebook: a measurement study of social network based applications, in: Proc. of the 8th ACM SIGCOMM
Conference on Internet Measurement, IMC ’08, ACM, New York, NY, USA, 2008, pp. 43–56.
[477] B. Viswanath, A. Mislove, M. Cha, K.P. Gummadi, On the evolution of user interaction in facebook, in: Proc. of the 2nd ACM workshop on Online social
network, ACM, New York, NY, USA, 2009, pp. 37–42.
[478] F. Schneider, A. Feldmann, B. Krishnamurthy, W. Willinger, Understanding online social network usage from a network perspective, in: Proc. of the
9th ACM SIGCOMM Conference on Internet Measurement Conference, IMC ’09, ACM, New York, NY, USA, 2009, pp. 35–48.
[479] N.P. Nguyen, T.N. Dinh, S. Tokala, M.T. Thai, Overlapping communities in dynamic networks: their detection and mobile applications, in: Proc. of the
17th Annual International Conference on Mobile Computing and Networking, MobiCom ’11, ACM, New York, NY, USA, 2011, pp. 85–96.
[480] H. Abdi, L.J. Williams, Principal component analysis, Wiley Interdiscip. Rev. Comput. Stat. 2 (4) (2010) 433–459.
[481] P. Kroonenberg, J. Leeuw, Principal component analysis of three-mode data by means of alternating least squares algorithms, Psychometrika 45 (1)
(1980) 69–97.
[482] D. Xu, S. Yan, L.Z. H., Zhang, Z. Liu, H.-Y. Shum, Concurrent subspaces analysis., in: CVPR (2), IEEE Computer Society, 2005, pp. 203–208.
[483] S.F. Everton, The noordin top terrorist network, in: Disrupting Dark Networks, Cambridge University Press, New York, 2012, pp. 385–397.
[484] W.D. Ivancic, Architecture and system engineering development study of space-based networks for nasa missions, in: Proc. IEEE Aerospace
Conference, ACM, 2003, pp. 1179–1186.
[485] R. Kali, J. Reyes, The architecture of globalization: a network approach to international economic integration, J. Int. Bus. Stud. 38 (4) (2007) 595–620.
[486] R.P. Keller, J.M. Drake, M.B. Drew, D.M. Lodge, Linking environmental conditions and ship movements to estimate invasive species transport across
the global shipping network, Diversity Distributions 17 (1) (2011) 93–102.
[487] C. Ducruet, T. Notteboom, The worldwide maritime network of container shipping: spatial structure and regional dynamics, Global Networks 12 (3)
(2012) 395–423.
[488] M. Boss, H. Elsinger, M. Summer, S. Thurner, Network topology of the interbank market, Quant. Finance 4 (6) (2004) 677–684.
[489] K. Soramäki, M.L. Bech, J. Arnold, R.J. Glass, W.E. Beyeler, The topology of interbank payment flows, Physica A 379 (1) (2007) 317–333.
[490] P. Gai, S. Kapadia, Contagion in financial networks, Proc. R. Soc. A 466 (2120) (2010) 2401–2423.
[491] L.M. Camarinha-Matos, H. Afsarmanesh, Collaborative networks: a new scientific discipline, J. Intell. Manuf. 16 (4–5) (2005) 439–452.
[492] L. Dooley, D. O’sullivan, Managing within distributed innovation networks, Int. J. Innov. Manag. 11 (03) (2007) 397–416.
[493] C. Durugbo, W. Hutabarat, A. Tiwari, J.R. Alcock, Modelling collaboration using complex networks, Inform. Sci. 181 (15) (2011) 3143–3161.
[494] M. Granovetter, Economic action and social structure: the problem of embeddedness, Am. J. Soc. (1985) 481–510.
[495] J. Galaskiewicz, W. Bielefeld, Nonprofit Organizations in an Age of Uncertainty: A Study of Organizational Change, Transaction Publishers, 1998.
[496] A.-L. Barabási, Z.N. Oltvai, Network biology: understanding the cell’s functional organization, Nat. Rev. Genet. 5 (2) (2004) 101–113.
[497] A.-L. Barabasi, N. Gulbahce, J. Loscalzo, Network medicine: a network-based approach to human disease, Nat. Rev. Genet. 12 (1) (2011) 56–68.
[498] H. Jeong, S.P. Mason, A.L. Barabási, Z.N. Oltvai, Lethality and centrality in protein networks, Nature 411 (6833) (2001) 41–42.
[499] R. Guimera, L.A.N. Amaral, Functional cartography of complex metabolic networks, Nature 433 (7028) (2005) 895–900.
[500] O. Sporns, The human connectome: a complex network, Ann. New York Acad. Sci. 1224 (1) (2011) 109–125.
[501] E. Bullmore, O. Sporns, Complex brain networks: graph theoretical analysis of structural and functional systems, Nat. Rev. Neurosci. 10 (3) (2009)
186–198.
[502] R.J. Fletcher, B.E. Bishop, R.P. Leon, R.A. Sclafani, C.M. Ogata, X.S. Chen, The structure and function of mcm from archaeal m. thermoautotrophicum,
Nat. Struct. Mol. Biol. 10 (3) (2003) 160–167.
[503] H. Kibak, L. Taiz, T. Starke, P. Bernasconi, J. Gogarten, Evolution of structure and function of v-atpases, J. Bioenerg. Biomemb. 24 (4) (1992) 415–424.
[504] K. Yamasaki, A. Gozolchiani, S. Havlin, Climate networks around the globe are significantly affected by el niño, Phys. Rev. Lett. 100 (2008) 228501.
[505] J.F. Donges, Y. Zou, N. Marwan, J. Kurths, The backbone of the climate network, Europhys. Lett. 87 (4) (2009) 48007.
[506] K. Steinhaeuser, A. Ganguly, N. Chawla, Multivariate and multiscale dependence in the global climate system revealed through complex networks,
Clim. Dynam. 39 (3–4) (2012) 889–895.
[507] L.H. de Saint-Denys, Dreams and How to Guide Them, Duckbacks, 1982.
[508] W. Güth, R. Schmittberger, B. Schwarze, An experimental analysis of ultimatum bargaining, J. Econ. Behav. Org. 3 (4) (1982) 367–388.
[509] G.E. Bolton, R. Zwick, Anonymity versus punishment in ultimatum bargaining, Games Econ. Behav. 10 (1) (1995) 95–121.
[510] M. Peltomäki, M. Alava, Three-and four-state rock–paper–scissors games with diffusion, Phys. Rev. E 78 (3) (2008) 031906.
[511] A. Baronchelli, M. Felici, V. Loreto, E. Caglioti, L. Steels, Sharp transition towards shared vocabularies in multi-agent systems, J. Stat. Mech. Theor.
Exp. 06 (2006) P06014. http://dx.doi.org/10.1088/1742-5468/2006/06/P06014.
[512] A. Baronchelli, L. Dall’Asta, A. Barrat, V. Loreto, Topology-induced coarsening in language games, Phys. Rev. E 73 (1) (2006) 015102.
[513] M. Milinski, R.D. Sommerfeld, H.-J. Krambeck, F.A. Reed, J. Marotzke, The collective-risk social dilemma and the prevention of simulated dangerous
climate change, Proc. Natl. Acad. Sci. USA 105 (7) (2008) 2291–2294.
[514] G. Hardin, The tragedy of the commons, Science 162 (3859) (1968) 1243–1248.
[515] C. Gracia-Lázaro, A. Ferrer, G. Ruiz, A. Tarancón, J.A. Cuesta, A. Sánchez, Y. Moreno, Heterogeneous networks do not promote cooperation when
humans play a prisoner’s dilemma, Proc. Natl. Acad. Sci. USA 109 (32) (2012) 12922–12926.
[516] S. Assenza, R. Gutiérrez, J. Gómez-Gardeñes, V. Latora, S. Boccaletti, Emergence of structural patterns out of synchronization in networks with
competitive interactions, Sci. Rep. 1, http://dx.doi.org/10.1038/srep00099.
[517] R. Gutiérrez, A. Amann, S. Assenza, J. Gómez-Gardeñes, V. Latora, S. Boccaletti, Emerging meso- and macroscales from synchronization of adaptive
networks, Phys. Rev. Lett. 107 (2011) 234103. http://dx.doi.org/10.1103/PhysRevLett.107.234103.
[518] V. Avalos-Gaytán, J.A. Almendral, D. Papo, S.E. Schaeffer, S. Boccaletti, Assortative and modular networks are shaped by adaptive synchronization
processes, Phys. Rev. E 86 (2012) 015101. http://dx.doi.org/10.1103/PhysRevE.86.015101.

Multi Capa

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Multi Capa

Uploaded by

Copyright:

Available Formats

Physics Reports ( ) –

Contents lists available at ScienceDirect

The structure and dynamics of multilayer networks

article info abstract

5.2.4. Voluntary prevention measures ................................................................................................................................. 70

1.1. The multilayer network approach to nature

1.2. Outline of the report

2. The structure of multilayer networks

2.1. Definitions and notations

2.1.1. The formal basic definitions

C = {Eαβ ⊆ Xα × Xβ ; α, β ∈ {1, . . . , M }, α ̸= β} (1)

The projection network of M is the graph proj(M ) = (XM , EM ) where

We will denote the adjacency matrix of proj(M ) = (XM , EM ) by AM .

(i) the links inside the different groups,

if aαij = 1 for some 1 ≤ α ≤ M

2.1.2. A small (and less formal) handbook

2.2. Characterizing the structure of multilayer networks

2.2.1. Centrality and ranking of nodes

where ki is the degree of the node i in the layer α , i.e. ki =

∀xαi , xαj ∈ Xα , c (xαi ) ∝ c (xαj ) if (xαj → xαi ) ∈ Gα , α ∈ {1, . . . , M }, (17)

So, the local heterogeneous eigenvector-like centrality matrix can be defined as

C⋆ = c⋆1 c⋆2 ... c⋆M ∈ R N ×M .

w11 (A[1] )T w12 (A[2] )T w1M (A[M ] )T

# of links between the neighbors of i

# of triangles in the network

Eα (i) = (k, j) ∈ Eα ; k, j ∈ Nα (i) .

Also, we have the relation

The following relationship between both clustering coefficients holds [36]:

|Nα∗ (i)|(|Nα∗ (i)| − 1) ly ly |Nα∗ (i)|(|Nα∗ (i)| − 1) ly

2.2.3. Metric structures: shortest paths and distances

where |XM | = N, and is also a way of measuring the performance of a graph.

be a path in M . The length of ω can be defined as the non-negative value

if ℓj ∈ C , (i.e. if ℓj is a crossed layer)

and β is an arbitrarily chosen non-negative parameter.

˜ (j) = β(Gσ (j−1) , Gσ (j) )

2.2.4. Matrices and spectral properties

λk ≤ µk ≤ λk+N −nα . (47)

2.2.5. Mesoscales: motifs and modular structures

2.3. Matrices and tensor representation: spectral properties

From node 1, we can access in one step nodes 2 and 3, since

{xi ⊗ ℓα ; i ∈ {1, . . . , N }, α ∈ {1, . . . , M }}. (61)

2.4. Correlations in multiplex networks

2.4.1. Degree correlations in multiplex networks

2.4.2. Overlap and multidegree

Thus, multiadjacency matrices satisfy the condition

⃗ have the explicit expression in terms of the adjacency

2.4.3. Multistrength and inverse multiparticipation ratio in weighted multiplex networks

2.4.4. The activities of the nodes and the pairwise multiplexity

PRE PRE PRE

k(m1,m2) k(m1,m2) k(m1,m2)

3.1. Multiplex networks models

3.1.1. Growing multiplex networks models

Πi[2] ∝ c2,1 k[i 1] + c2,2 k[i 2] , (85)

2Γ (2 + 2m)Γ (k)Γ (q)Γ (k + q − 2m + 1)

Therefore, the two layer have positive degree correlations.

3.1.2. Simple static models

3.1.3. Ensembles of multiplex networks

⟨aαij aαij ⟩ = ⟨aαij ⟩⟨aαij ⟩

⟨aαij aαij ⟩ ̸= ⟨aαij ⟩⟨aαij ⟩.

• (ii) Fixed multidegree sequence

where P (Gα |{⃗

3.2. Ensembles of networks of networks

k̄αi k̄αj k̄αi k̄αj

3.2.2. Case II: networks of networks with given superdegree distribution